KO | EN

GitLite

Minimalist Open Source Repository Search Engine

gitlite — search

#dsh #dsh-plugin #developer-tools #ai-agent #claude-code #codex #typescript #ai-agents #deepseek-harness #mcp #cli #rust

Awesome-Video-Action-Model

★ 75 Open GitHub ↗

No description available.

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# 🎬 Awesome-Video-Action-Models

Video action models grounded in video prediction, video generation, and world-action modeling demonstrate strong potential in embodied control. They can serve as video-conditioned policies, joint action-world generators, unified foundation models that support mixed training with action-labeled trajectories and action-free videos, and bridge mechanisms from human or action-free video to executable robot behavior.
This repository summarizes recent research on **video action models (VAM)**, **world-action models (WAM)**, and **video-to-robot bridge signals** in embodied AI.

## 🗂️ Quick Navigation

- 🧭 [Survey / Position Papers](#survey-position-papers)
- 1️⃣ [Route 1: Two-Stage Video Prior -> Action](#route-1-two-stage-video-prior-to-action)
- 2️⃣ [Route 2: Joint Video-Action Generation](#route-2-joint-video-action-generation)
- 3️⃣ [Route 3: Unified Video-Action / World-Action Models](#route-3-unified-video-action-world-action-models)
- 4️⃣ [Route 4: Physical Intelligence, Bridge-Heavy Signals, and Evaluation](#route-4-physical-intelligence-bridge-heavy-signals-and-evaluation)
- 🔀 [Bridge Signal View](#bridge-signal-view)
- 🕹️ [VAM-RL View](#vam-rl-view)

<a id="survey-position-papers"></a>
# 🧭 Survey / Position Papers

These papers are not single VAM/WAM methods, but they are useful anchors for framing the field and organizing the route taxonomy below.

| Paper | Venue | Date | Scope | Code/Project | Why it matters |
| --- | --- | --- | --- | --- | --- |
| [World Action Models: The Next Frontier in Embodied AI](https://arxiv.org/abs/2605.12090) | arXiv 2026 | 5/2026 | `position; WAM taxonomy` | [Project](https://openmoss.github.io/Awesome-WAM) / [Code](https://github.com/OpenMOSS/Awesome-WAM) | Defines WAMs as a central embodied-AI interface and helps motivate the split between explicit video imagination, latent future modeling, action-only deployment, and evaluation. |
| [World Model for Robot Learning: A Comprehensive Survey](https://arxiv.org/abs/2605.00080) | arXiv 2026 | 4/2026 | `survey; robot world models` | [Project](https://ntumars.github.io/wm-robot-survey/) / [Code](https://github.com/NTUMARS/Awesome-World-Model-for-Robotics-Policy) | Broader robot world-model map, useful as an external survey reference beyond VAM/WAM-specific methods. |
| [From Pixels to Tokens](https://arxiv.org/abs/2605.04678) | arXiv 2026 | 5/2026 | `systematic study; latent action supervision` | [Code](https://github.com/RUCKBReasoning/From_Pixels_to_Tokens) | Controlled study of image-based and action-based latent action supervision, useful for deciding when latent tokens should be used as alignment targets, explicit plans, conditional prefixes, or action-token objectives. |
| [Physically Native World Models](https://arxiv.org/abs/2605.00412) | arXiv 2026 | 5/2026 | `position; physics-native world models` | - | Position/theory anchor arguing that embodied world models should move beyond visual plausibility toward physically structured latent dynamics, phase-space state, controllability, and long-horizon rollout stability. |
| [World Models for Robotic Manipulation: A Survey](https://arxiv.org/abs/2606.00113) | arXiv 2026 | 6/2026 | `survey; robot world models` | - | Survey anchor that clarifies robot-world-model boundaries, taxonomy, and functional roles across latent dynamics, video rollouts, learned simulators, policy-facing planning, and evaluation. |
| [Cosmos 3](https://arxiv.org/abs/2606.02800) | arXiv 2026 | 6/2026 | `system; physical AI world model` | [Project](https://research.nvidia.com/labs/cosmos-lab/cosmos3) / [Code](https://github.com/NVIDIA/cosmos) / [Ckpt](https://huggingface.co/collections/nvidia/cosmos3) | Omnimodal physical-AI world-model system spanning language, image, video, audio, and action; best treated as a large-system reference rather than a single robot VAM method. |
| [From World Models to World Action Models](https://arxiv.org/abs/2607.00836) | arXiv 2026 | 7/2026 | `tutorial; WAM taxonomy` | [Project](https://clearlab-sustech.github.io/WorldModelSurvey/) / [Code](https://github.com/clearlab-sustech/WorldModelSurvey) | Concise tutorial that turns world-model design choices into a WAM-oriented taxonomy, giving the route split below a cleaner conceptual anchor. |

<a id="route-1-two-stage-video-prior-to-action"></a>
# 1️⃣ Route 1: Two-Stage Video Prior -> Action

1️⃣ `R1` is where video or world modeling comes first and executable action is recovered second.

| Method | Title | Venue | Date | Data tags | Code/Project | Key feature/finding |
| --- | --- | --- | --- | --- | --- | --- |
| [Dreamitate](https://arxiv.org/abs/2406.16862) | Dreamitate: Real-World Visuomotor Policy Learning via Video Generation | CoRL 2024 | 6/2024 | `robot-only; exo; action-labeled; scripted` | [Project](https://dreamitate.cs.columbia.edu/) / [Code](https://github.com/cvlab-columbia/dreamitate) | Bridge: `World Model / Video Plan`. Strict VAM: `Partial`. Uses video generation as a policy interface and decodes tool motion into robot execution. |
| [Gen2Act](https://arxiv.org/abs/2409.16283) | Gen2Act: Human Video Generation in Novel Scenarios enables Generalizable Robot Manipulation | arXiv 2024 | 9/2024 | `co-training; exo; action-free + action-labeled; scripted` | [Project](https://homangab.github.io/gen2act/) | Bridge: `World Model / Human Video Plan`. Strict VAM: `Yes`. Generates human task videos zero-shot and translates them into closed-loop robot actions through a video-conditioned policy. |
| [UniPi](https://arxiv.org/abs/2302.00111) | Learning Universal Policies via Text-Guided Video Generation | NeurIPS 2023 | 2/2023 | `robot-only; exo; action-labeled; scripted` | [Code](https://github.com/universal-policy/unipi) | Bridge: `World Model / Video Plan`. Strict VAM: `Partial`. Canonical policy-as-video formulation: generate a future plan in image space, then infer action. |
| [GR-1](https://arxiv.org/abs/2312.13139) | Unleashing Large-Scale Video Generative Pre-training for Visual Robot Manipulation | ICLR 2024 | 12/2023 | `robot-only; exo; action-labeled; scripted` | [Project](https://gr1-manipulation.github.io/) / [Code](https://github.com/bytedance/GR-1) | Bridge: `World Model / Future Image Prediction`. Strict VAM: `Partial`. Early anchor that couples large-scale video generative pretraining with action prediction and future image prediction in one robot manipulation transformer. |
| [VILP](https://arxiv.org/abs/2502.01784) | VILP: Imitation Learning with Latent Video Planning | arXiv 2025 | 2/2025 | `robot-only; multi-view; action-labeled; scripted` | [Code](https://github.com/ZhengtongXu/VILP) | Bridge: `World Model / Video Plan`. Strict VAM: `Yes`. Uses latent video diffusion as a high-level planner and a low-level video-to-action policy, then replans in receding horizon for closed-loop control. |
| [Grounding Video Models to Actions through Goal Conditioned Exploration](https://arxiv.org/abs/2411.07223) | Grounding Video Models to Actions through Goal Conditioned Exploration | ICLR 2025 Spotlight | 11/2024 | `robot-only; exo; action-free + action-labeled; scripted` | - | Bridge: `World Model / Visual Goal`. Strict VAM: `Partial`. Turns generated future frames into reachable visual subgoals and grounds them with self-supervised goal-conditioned control. |
| [LAPA](https://arxiv.org/abs/2410.11758) | Latent Action Pretraining from Videos | ICLR 2025 | 10/2024 | `co-training; exo; action-free + action-labeled; scripted` | [Project](https://latentactionpretraining.github.io/) / [Code](https://github.com/LatentActionPretraining/LAPA) / [Ckpt](https://huggingface.co/latent-action-pretraining/LAPA-7B-openx) | Bridge: `Latent Action`. Strict VAM: `No`. Important pretraining route for extracting pseudo-actions from video before robot grounding. |
| [VPP](https://arxiv.org/abs/2412.14803) | Video Prediction Policy: A Generalist Robot Policy with Predictive Visual Representations | ICML 2025 Spotlight | 12/2024 | `robot-only; exo; action-labeled; scripted` | [Project](https://video-prediction-policy.github.io/) / [Code](https://github.com/roboterax/video-prediction-policy) / [Ckpt](https://huggingface.co/yjguo/svd-robot-calvin-ft) | Bridge: `World Model / Predictive Representation`. Strict VAM: `Yes`. Predictive video features are used directly as the control representation. |
| [Seer](https://arxiv.org/abs/2412.15109) | Predictive Inverse Dynamics Models are Scalable Learners for Robotic Manipulation | ICLR 2025 | 12/2024 | `robot-only; exo; action-labeled; scripted` | [Project](https://nimolty.github.io/Seer/) / [Code](https://github.com/InternRobotics/Seer) | Bridge: `World Model / Predictive Visual State`. Strict VAM: `Partial`. Forecasts future visual states and recovers actions through predictive inverse dynamics, making it a clean staged predictive-control anchor. |
| [FLARE](https://arxiv.org/abs/2505.15659) | FLARE: Robot Learning with Implicit World Modeling | CoRL 2025 | 5/2025 | `co-training; ego; action-free + action-labeled; scripted` | [Project](https://research.nvidia.com/labs/gear/flare/) | Bridge: `World Model / Future Latent`. Strict VAM: `Partial`. Aligns policy tokens with latent future observations so an action model can reason about long-term consequences without explicit video rollout. |
| [LaDi-WM](https://arxiv.org/abs/2505.11528) | LaDi-WM: A Latent Diffusion-based World Model for Predictive Manipulation | CoRL 2025 | 5/2025 | `robot-only; exo; action-labeled; scripted` | [Project](https://guhuangai.github.io/LaDiWM.github.io/) / [Code](https://github.com/GuHuangAI/LaDiWM) | Bridge: `World Model / Future Latent`. Strict VAM: `Partial`. Uses a VFM-aligned latent diffusion world model to imagine future states and iteratively refine actions for predictive manipulation. |
| [FRAPPE](https://arxiv.org/abs/2602.17259) | FRAPPE: Infusing World Modeling into Generalist Policies via Multiple Future Representation Alignment | arXiv 2026 | 2/2026 | `co-training; exo; action-free + action-labeled; scripted` | [Project](https://h-zhao1997.github.io/frappe) / [Code](https://github.com/OpenHelix-Team/frappe) / [Ckpt](https://huggingface.co/collections/hhhJB/frappe) | Bridge: `World Model / Future Representation`. Strict VAM: `Partial`. Support: `Yes`. Aligns multiple future visual representations into a generalist policy during training while keeping deployment action-only and free of rollout feedback. |
| [DreamGen](https://arxiv.org/abs/2505.12705) | DreamGen: Unlocking Generalization in Robot Learning through Video World Models | arXiv 2025 | 5/2025 | `robot-only; exo; action-labeled; scripted` | - | Bridge: `World Model / Synthetic Neural Trajectory`. Strict VAM: `No`. Support: `Yes`. Uses a video world model plus pseudo-action recovery to synthesize neural trajectories for downstream robot policies. |
| [ConLA](https://arxiv.org/abs/2602.00557) | ConLA: Contrastive Latent Action Learning from Human Videos for Robotic Manipulation | arXiv 2026 | 2/2026 | `co-training; exo; action-free + action-labeled; scripted` | [Code](https://github.com/WeishengDAI/ConLA) | Bridge: `Latent Action`. Strict VAM: `No`. Strengthens latent-action pretraining with contrastive token shaping so human videos transfer more cleanly into robot control. |
| [DreamDojo](https://arxiv.org/abs/2602.06949) | DreamDojo: A Generalist Robot World Model from Large-Scale Human Videos | arXiv 2026 | 2/2026 | `co-training; ego; action-free + action-labeled; scripted` | [Project](https://dreamdojo-world.github.io/) | Bridge: `Latent Action`. Strict VAM: `Partial`. Support: `Yes`. Uses continuous latent actions to turn large-scale egocentric human video into action-conditioned robot world-model pretraining, then distills the result into a streamable interactive model. |
| [Say, Dream, and Act](https://arxiv.org/abs/2602.10717) | Say, Dream, and Act: Learning Video World Models for Instruction-Driven Robot Manipulation | arXiv 2026 | 2/2026 | `robot-only; exo; action-labeled; scripted` | - | Bridge: `World Model / Video Plan`. Strict VAM: `Yes`. Selects, adapts, and distills a video world model, then conditions action generation on imagined keyframes plus real observation history. |
| [DiT4DiT](https://arxiv.org/abs/2603.10448) | DiT4DiT: Jointly Modeling Video Dynamics and Actions for Generalizable Robot Control | arXiv 2026 | 3/2026 | `robot-only; ego; action-labeled; scripted` | [Project](https://dit4dit.github.io/) | Bridge: `World Model / Predictive Representation`. Strict VAM: `Yes`. Cascades a video DiT into an action DiT through intermediate denoising features instead of rendered future frames. |
| [FutureVLA](https://arxiv.org/abs/2603.10712) | FutureVLA: Joint Visuomotor Prediction for Vision-Language-Action Model | arXiv 2026 | 3/2026 | `robot-only; exo; action-labeled; scripted` | [Project](https://xuxiaoxxxx.github.io/projects/futurevla/) / [Code](https://github.com/xuxiaoxxxx/FutureVLA) | Bridge: `World Model / Joint Visuomotor Prior`. Strict VAM: `Partial`. Learns a continuous joint visuomotor predictive prior and aligns it into a downstream VLA without changing the deployed action head. |
| [DIAL](https://arxiv.org/abs/2603.29844) | DIAL: Decoupling Intent and Action via Latent World Modeling for End-to-End VLA | arXiv 2026 | 3/2026 | `co-training; exo; action-free + action-labeled; scripted` | - | Bridge: `World Model / Latent Foresight`. Strict VAM: `Yes`. Makes latent future intent a mandatory control bottleneck by coupling a predictive System-2 prior with a latent inverse-dynamics actor. |
| [S-VAM](https://arxiv.org/abs/2603.16195) | S-VAM: Shortcut Video-Action Model by Self-Distilling Geometric and Semantic Foresight | arXiv 2026 | 3/2026 | `robot-only; exo; action-labeled; scripted` | [Project](https://haodong-yan.github.io/S-VAM/) / [Code](https://github.com/Haodong-Yan/S-VAM-Code) | Bridge: `World Model / Predictive Representation`. Strict VAM: `Yes`. Self-distills multi-step video priors into one-step geometric and semantic foresight for fast closed-loop control. |
| [StructVLA](https://arxiv.org/abs/2603.12553) | Beyond Dense Futures: World Models as Structured Planners for Robotic Manipulation | arXiv 2026 | 3/2026 | `robot-only; exo; action-labeled; scripted` | - | Bridge: `World Model / Structured Foresight`. Strict VAM: `Partial`. Replaces dense future rollouts with kinematically grounded structured frames and then maps sparse foresight into low-level control. |
| [World2Act](https://arxiv.org/abs/2603.10422) | World2Act: Latent Action Post-Training via Skill-Compositional World Models | arXiv 2026 | 3/2026 | `robot-only; exo; action-labeled; scripted` | [Project](https://wm2act.github.io/) | Bridge: `Latent Action / World-Model Dynamics Matching`. Strict VAM: `No`. Support: `Yes`. Post-trains a residual policy by aligning action chunks with world-model video-dynamics latents instead of pixel-space rollouts. |
| [VAMPO](https://arxiv.org/abs/2603.19370) | VAMPO: Policy Optimization for Improving Visual Dynamics in Video Action Models | arXiv 2026 | 3/2026 | `robot-only; exo; action-labeled; scripted` | [Project](https://vampo-robot.github.io/VAMPO/) / [Code](https://github.com/OpenHelix-Team/VAMPO) / [Ckpt](https://huggingface.co/williammmgezju/VAMPO_SVD) | Bridge: `World Model / Predictive Representation`. Strict VAM: `Yes`. Optimization tag: `VAM-RL`. Post-trains a VPP-style video predictor with GRPO and Euler Hybrid sampling to improve control-relevant visual dynamics. |
| [Veo-Act](https://arxiv.org/abs/2604.04502) | Veo-Act: How Far Can Frontier Video Models Advance Generalizable Robot Manipulation? | arXiv 2026 | 4/2026 | `robot-only; multi-view; action-labeled; scripted` | - | Bridge: `World Model / Video Plan`. Strict VAM: `Partial`. Uses Veo-3 as a high-level motion planner and switches to a VLA low-level executor via an IDM-based interaction detector for dexterous manipulation. |
| [StableIDM](https://arxiv.org/abs/2604.17887) | StableIDM: Stabilizing Inverse Dynamics Model against Manipulator Truncation via Spatio-Temporal Refinement | arXiv 2026 | 4/2026 | `robot-only; exo; action-labeled; scripted` | - | Bridge: `World Model / Inverse Dynamics Action Recovery`. Strict VAM: `No`. Stabilizes video-to-action inverse dynamics under manipulator truncation with robot-centric masking, directional aggregation, and temporal refinement, making staged VGM-to-action pipelines less brittle. |
| [Mask World Model](https://arxiv.org/abs/2604.19683) | Mask World Model: Predicting What Matters for Robust Robot Policy Learning | arXiv 2026 | 4/2026 | `robot-only; multi-view; action-labeled; scripted` | [Code](https://github.com/LYFCLOUDFAN/mask-world-model) | Bridge: `Explicit 2D / Future Semantic Mask`. Strict VAM: `Partial`. Predicts future semantic-mask dynamics instead of RGB and feeds mask-centric future features into policy learning, forcing foresight to focus on task-relevant geometry, contact, and object motion. |
| [T3VF](https://arxiv.org/abs/2605.08215) | Test-Time Training for Visual Foresight Vision-Language-Action Models | arXiv 2026 | 5/2026 | `robot-only; exo; action-labeled; scripted` | - | Bridge: `World Model / Visual Foresight Self-Supervision`. Strict VAM: `No`. Support: `Yes`. Uses predicted-attained future observation pairs and action-variance filtering to test-time update visual-foresight paths for deployed VLA policies. |
| [MoLA](https://arxiv.org/abs/2605.12167) | From Imagined Futures to Executable Actions: Mixture of Latent Actions for Robot Manipulation | ICML 2026 | 5/2026 | `co-training; exo; action-free + action-labeled; scripted` | [Project](https://logosroboticsgroup.github.io/MoLA/) / [Code](https://github.com/LogosRoboticsGroup/MoLA) | Bridge: `Latent Action / Future-to-Action IDM`. Strict VAM: `Partial`. Inverts imagined semantic, depth, and flow futures into latent actions, then generates executable robot actions through a flow action head. |
| [mimic-video](https://arxiv.org/abs/2512.15692) | mimic-video: Video-Action Models for Generalizable Robot Control Beyond VLAs | arXiv 2025 | 12/2025 | `robot-only; exo; action-labeled; scripted` | [Project](https://mimic-video.github.io/) | Bridge: `World Model / Predictive Representation`. Strict VAM: `Yes`. Freezes a video backbone and controls through partial-denoising latent interfaces. |
| [Vidar](https://arxiv.org/abs/2507.12898) | Vidar: Embodied Video Diffusion Model for Generalist Manipulation | arXiv 2025 | 7/2025 | `robot-only; exo; action-labeled; scripted` | [Project](https://embodiedfoundation.github.io/vidar_anypos) / [Code](https://github.com/thu-ml/vidar) / [Ckpt](https://huggingface.co/Xiang-cd/vidar) | Bridge: `World Model / Predictive Representation`. Strict VAM: `Partial`. Strong video prior plus lightweight embodiment alignment. |
| [Vidarc](https://arxiv.org/abs/2512.17661) | Vidarc: Embodied Video Diffusion Model for Closed-loop Control | arXiv 2025 | 12/2025 | `robot-only; exo; action-labeled; scripted` | [Project](https://embodiedfoundation.github.io/vidar_anypos) / [Code](https://github.com/thu-ml/vidar) / [Ckpt](https://huggingface.co/Xiang-cd/vidar) | Bridge: `World Model / Predictive Representation`. Strict VAM: `Partial`. Closed-loop extension of embodied video diffusion with feedback refill and lower-latency control. |
| [TC-IDM](https://arxiv.org/abs/2601.18323) | TC-IDM: Grounding Video Generation for Executable Zero-shot Robot Motion | arXiv 2026 | 1/2026 | `robot-only; exo; action-labeled; scripted` | [Project](https://wsbaiyi.github.io/TC-IDM-Page/) / [Code](https://github.com/wsbaiyi/TC-IDM) | Bridge: `World Model / Video Plan`. Strict VAM: `Partial`. Grounds generated motion into executable zero-shot robot behavior through action recovery. |
| [DVA](https://www.rhoda.ai/research/direct-video-action) | Causal Video Models Are Data-Efficient Robot Policy Learners | Rhoda AI Research 2026 | 3/2026 | `co-training; exo; action-free + action-labeled; unscripted` | [Project](https://www.rhoda.ai/research/direct-video-action) | Bridge: `World Model / Video Plan`. Strict VAM: `Yes`. Recasts policy inference as real-time causal future-video prediction, then uses an inverse-dynamics model and leapfrog inference to recover executable actions. |
| [VERA](https://arxiv.org/abs/2605.27817) | Turning Video Models into Generalist Robot Policies | arXiv 2026 | 5/2026 | `co-training; exo; action-free + action-labeled; unscripted` | [Project](https://vera.csail.mit.edu) | Bridge: `World Model / Video Plan + Jacobian IDM`. Strict VAM: `Yes`. Keeps an action-free video model as the task-level visual planner and grounds imagined flow changes into embodiment-specific robot actions through a Jacobian inverse-dynamics model. |
| [Demo-JEPA](https://arxiv.org/abs/2605.20811) | Demo-JEPA: Joint-Embedding Predictive Architecture for One-shot Cross-Embodiment Imitation | arXiv 2026 | 5/2026 | `co-training; exo; action-free + action-labeled; scripted` | [Project](https://log2r.github.io/Demo-JEPA/) / [Code](https://github.com/banban3forever/Demo-JEPA) | Bridge: `World Model / Latent Goal`. Strict VAM: `Partial`. Converts a source embodiment demonstration into a target-compatible JEPA latent future goal, then plans under the target robot's learned dynamics. |
| [SeeTraceAct](https://arxiv.org/abs/2606.02745) | SeeTraceAct: Visibility-Aware Latent Planning from Cross-Embodiment Demonstration Videos | arXiv 2026 | 6/2026 | `co-training; exo; action-free + action-labeled; scripted` | [Code](https://github.com/jaehyeon-son/SeeTraceAct) | Bridge: `World Model / Visibility-Aware Latent Trace`. Strict VAM: `Partial`. Boundary bridge from cross-embodiment demonstration videos to visible future traces and target-robot action generation. |
| [omega-EVA](https://arxiv.org/abs/2606.09457) | omega-EVA: Envision, Verify, and Act with Latent Interactive World Models | arXiv 2026 | 6/2026 | `robot-only; exo; action-labeled; scripted` | - | Bridge: `World Model / Candidate-Action Latent Future`. Strict VAM: `Partial`. Uses a latent interactive world model to check candidate action consequences before refinement, making it an R1/R4 boundary method rather than a plain VLA. |
| [MemoryVLA++](https://arxiv.org/abs/2606.09827) | MemoryVLA++: Temporal Modeling via Memory and Imagination in Vision-Language-Action Models | arXiv 2026 | 6/2026 | `robot-only; exo; action-labeled; scripted` | - | Bridge: `World Model / Memory-Guided Latent Imagination`. Strict VAM: `Partial`. Boundary case: injects video-diffusion latent imagination and memory into a VLA action path, but is not a canonical standalone WAM. |
| [LaWAM](https://arxiv.org/abs/2606.15768) | LaWAM: Latent World Action Models for Efficient Dynamics-Aware Robot Policies | arXiv 2026 | 6/2026 | `co-training; ego; action-free + action-labeled; scripted` | [Project](https://rlinf.github.io/LaWAM/) / [Code](https://github.com/RLinf/LaWAM) / [Ckpt](https://huggingface.co/collections/jialei02/lawam-checkpoints) | Bridge: `World Model / Latent Visual Subgoal`. Strict VAM: `Yes`. Learns a latent-action-conditioned world model first, then uses predicted latent visual subgoals to condition chunk-level action generation with low-latency inference. |
| [ImageWAM](https://arxiv.org/abs/2606.19531) | ImageWAM: Do World Action Models Really Need Video Generation, or Just Image Editing? | arXiv 2026 | 6/2026 | `robot-only; exo; action-labeled; scripted` | [Project](https://zhangwenyao1.github.io/ImageWAM/) / [Code](https://github.com/yuyangalin/ImageWAM) / [Ckpt](https://huggingface.co/collections/yuyangalin/imagewam) | Bridge: `World Model / Image-Editing Cache`. Strict VAM: `Partial`. Replaces dense future-video rollout with image-editing denoising caches, then conditions a flow-matching action expert on the compact change-centric world-action context. |
| [MemoryVAM](https://arxiv.org/abs/2606.20679) | MemoryVAM: Integrating Memory into Video Action Model for Robot Manipulation | arXiv 2026 | 6/2026 | `robot-only; exo; action-labeled; scripted` | [Project](https://memoryvam.github.io/) | Bridge: `World Model / Memory-Conditioned Video Plan`. Strict VAM: `Partial`. Support: `Yes`. Adds episodic Recap-Cue memory to video-action models so episode history conditions both future prediction and action decoding for long-horizon manipulation. |
| [DVG-WM](https://arxiv.org/abs/2606.32028) | DVG-WM: Disentangled Video Generation Enables Efficient Embodied World Model for Robotic Manipulation | ECCV 2026 | 6/2026 | `robot-only; multi-view; action-labeled; scripted` | [Project](https://zyshan0929.github.io/DVGWM/) / [Code](https://github.com/zyshan0929/DVG-WM) | Bridge: `World Model / Disentangled Video Plan`. Strict VAM: `Partial`. Splits future generation into low-resolution dynamics preview and high-resolution refinement, then feeds predicted futures to a vision-only diffusion policy for efficient video-conditioned control. |
| [VLAFlow](https://arxiv.org/abs/2607.01586) | VLAFlow: A Unified Training Framework for Vision-Language-Action Models via Co-training and Future Latent Alignment | arXiv 2026 | 7/2026 | `robot-only; exo; action-labeled; scripted` | [Project](https://mindvla-team.github.io/VLAFlow/) / [Code](https://github.com/MindVLA-Team/VLAFlow) | Bridge: `World Model / Future Latent Alignment`. Strict VAM: `No`. Support: `Yes`. Uses frozen future-latent prediction as a controlled VLA training signal, improving action supervision without requiring test-time video rollout. |
| [KAM-WM](https://arxiv.org/abs/2607.04652) | KAM-WM: Kinematic Affordance Maps from Latent World Models for Robot Manipulation | arXiv 2026 | 7/2026 | `robot-only; exo; action-labeled; scripted` | - | Bridge: `Explicit 2D / Kinematic Affordance Map`. Strict VAM: `Partial`. Extracts directional kinematic affordances from the latent velocity of a frozen image-to-video world model, then conditions a diffusion policy on that structured motion interface. |
| [RoboTALES](https://arxiv.org/abs/2607.06018) | RoboTALES: Learning Reasoning-Guided Robot Policies via Task-Aligned Simulated Futures | arXiv 2026 | 7/2026 | `robot-only; exo; action-labeled; scripted` | [Code](https://github.com/hananshafi/RoboTALES) | Bridge: `World Model / Task-Aligned Simulated Future`. Strict VAM: `Partial`. Uses language-planned subgoals and critic feedback to make generated futures task-aligned before grounding their representations into a robot policy. |
| [GenVid2Robot](https://arxiv.org/abs/2607.09191) | GenVid2Robot: From Video Generation to Robot Manipulation via Rigid-Geometric Consistency | arXiv 2026 | 7/2026 | `robot-only; exo RGB-D; action-labeled; scripted` | - | Bridge: `Explicit 3D / Video Anchors + SE(3) Trajectory`. Strict VAM: `Partial`. Tracks semantic anchors in generated video, validates them with RGB-D rigid-geometric consistency, and converts the resulting relative motion into grasp-conditioned executable trajectories. |
| [Jetson-PI](https://arxiv.org/abs/2607.12659) | Jetson-PI: Towards Onboard Real-Time Robot Control via Foresight-Aligned Asynchronous Inference | arXiv 2026 | 7/2026 | `robot-only; exo; action-labeled; scripted` | [Code](https://github.com/PKU-SEC-Lab/Jetson-PI) | Bridge: `World Model / Future Representation Correction`. Strict VAM: `Partial`. Predicts the future environment representation induced by already committed actions, then lets the action expert operate from the corrected future time step for asynchronous onboard control. |
| [RxBrain](https://arxiv.org/abs/2607.14187) | RxBrain: Embodied Cognition Foundation Model with Joint Language-Visual Reasoning and Imagination | arXiv 2026 | 7/2026 | `robot-only; exo; action-labeled; scripted` | - | Bridge: `World Model / Visual State Imagination + Subgoal`. Strict VAM: `Partial`. Unifies language planning, visual world-state imagination, subgoal generation, and continuous robot action inside one embodied-cognition model. |
| [FoMoVLA](https://arxiv.org/abs/2607.14739) | FoMoVLA: Bridging Visual Foresight and Motion Guidance for Vision-Language-Action Models | arXiv 2026 | 7/2026 | `robot-only; exo; action-labeled; scripted` | [Project](https://liauto-research.github.io/FoMoVLA/) | Bridge: `Explicit 2D / Future Feature + Point Tracks`. Strict VAM: `Partial`. Jointly predicts future feature states and sparse 2D point trajectories, then couples both signals to the continuous action policy through future-conditioned cross-attention. |
| [Masked Visual Actions](https://arxiv.org/abs/2607.19343) | Masked Visual Actions for Unified World Modeling | arXiv 2026 | 7/2026 | `robot-only; exo; action-labeled; scripted` | [Project](https://masked-visual-actions.github.io/) / [Code](https://github.com/HadiZayer/masked-visual-actions) | Bridge: `Explicit 2D / Masked Entity Trajectory`. Strict VAM: `Partial`. Treats partially revealed pixel-space entity trajectories as actions, supporting forward prediction, inverse action recovery, planning, and policy evaluation in one visual interface. |
| [HOST](https://arxiv.org/abs/2607.20033) | Robots Acquire Manipulation Skills in Seconds from a Single Human Video | arXiv 2026 | 7/2026 | `co-training; exo; action-free + action-labeled; scripted` | - | Bridge: `World Model / Human Video -> Robot Future Observation`. Strict VAM: `Yes`. Converts a single human demonstration into task progress, robot future observations, and finally executable actions, making predicted robot futures a mandatory control bottleneck. |

<a id="route-2-joint-video-action-generation"></a>
# 2️⃣ Route 2: Joint Video-Action Generation

2️⃣ `R2` is where action and future consequence are learned inside one coupled generation objective.

| Method | Title | Venue | Date | Data tags | Code/Project | Key feature/finding |
| --- | --- | --- | --- | --- | --- | --- |
| [PAD](https://arxiv.org/abs/2411.18179) | Prediction with Action: Visual Policy Learning via Joint Denoising Process | NeurIPS 2024 | 11/2024 | `co-training; exo; action-free + action-labeled; scripted` | [Code](https://github.com/Robert-gyj/Prediction_with_Action) / [Ckpt](https://huggingface.co/yjguo/pad_bridge_pre) | Bridge: `World Model / Joint Video-Action`. Strict VAM: `Partial`. Clean joint-denoising baseline for future consequence and action coupling. |
| [GR-2](https://arxiv.org/abs/2410.06158) | GR-2: A Generative Video-Language-Action Model with Web-Scale Knowledge for Robot Manipulation | arXiv 2024 | 10/2024 | `co-training; exo; action-free + action-labeled; scripted` | - | Bridge: `World Model / Joint Video-Action`. Strict VAM: `Partial`. Couples future video prediction with action-trajectory generation after web-scale pretraining, making it an early generative VLA/WAM anchor. |
| [VideoVLA](https://arxiv.org/abs/2512.06963) | VideoVLA: Video Generators Can Be Generalizable Robot Manipulators | NeurIPS 2025 | 12/2025 | `robot-only; exo; action-labeled; scripted` | [Project](https://videovla-nips2025.github.io/) / [Code](https://github.com/VideoVLA-Project/VideoVLA) | Bridge: `World Model / Joint Video-Action`. Strict VAM: `Partial`. Explicitly argues that a video generator can be turned into a generalizable manipulator. |
| [CoVAR](https://arxiv.org/abs/2512.16023) | CoVAR: Co-generation of Video and Action for Robotic Manipulation via Multi-Modal Diffusion | arXiv 2025 | 12/2025 | `robot-only; exo; action-labeled; scripted` | - | Bridge: `World Model / Joint Video-Action`. Strict VAM: `Yes`. Uses parallel video and action diffusion models with bridge attention so pretrained video knowledge is preserved while video-action pairs are co-generated. |
| [Cosmos Policy](https://arxiv.org/abs/2601.16163) | Cosmos Policy: Fine-Tuning Video Models for Visuomotor Control and Planning | arXiv 2026 | 1/2026 | `robot-only; exo; action-labeled; scripted` | [Project](https://research.nvidia.com/labs/cosmos-lab/cosmos-policy/) / [Code](https://github.com/NVlabs/cosmos-policy) / [Ckpt](https://huggingface.co/collections/nvidia/cosmos-policy) | Bridge: `World Model / Joint Video-Action`. Strict VAM: `Partial`. Treats actions, future states, and value as latent frames in one video-model pipeline. |
| [DreamZero](https://arxiv.org/abs/2602.15922) | World Action Models are Zero-shot Policies | arXiv 2026 | 2/2026 | `robot-only; exo; action-labeled; scripted` | [Project](https://dreamzero0.github.io/) / [Code](https://github.com/dreamzero0/DreamZero) / [Ckpt](https://huggingface.co/GEAR-Dreams/DreamZero-DROID) | Bridge: `World Model / Joint Video-Action`. Strict VAM: `Partial`. Systems-heavy WAM for scaling joint action-world modeling into real closed-loop use. |
| [WorldVLA](https://arxiv.org/abs/2506.21539) | WorldVLA: Towards Autoregressive Action World Model | arXiv 2025 | 6/2025 | `robot-only; exo; action-labeled; scripted` | [Code](https://github.com/alibaba-damo-academy/RynnVLA-002) / [Ckpt](https://huggingface.co/Alibaba-DAMO-Academy/WorldVLA) | Bridge: `World Model / Joint Video-Action`. Strict VAM: `Partial`. Autoregressive action-world modeling with unified token space. |
| [Video Generators are Robot Policies](https://arxiv.org/abs/2508.00795) | Video Generators are Robot Policies | arXiv 2025 | 8/2025 | `co-training; exo; action-free + action-labeled; scripted` | [Code](https://github.com/cvlab-columbia/videopolicy) | Bridge: `World Model / Joint Video-Action`. Strict VAM: `Partial`. Jointly denoises future video and action while treating the video generator as the primary policy backbone and the action branch as its executable interface. |
| [LingBot-VA](https://arxiv.org/abs/2601.21998) | Causal World Modeling for Robot Control | arXiv 2026 | 1/2026 | `robot-only; exo; action-labeled; scripted` | [Project](https://technology.robbyant.com/lingbot-va) / [Code](https://github.com/robbyant/lingbot-va) / [Ckpt](https://huggingface.co/collections/robbyant/lingbot-va) | Bridge: `World Model / Joint Video-Action`. Strict VAM: `Partial`. Causal autoregressive world modeling plus closed-loop refill and asynchronous inference. |
| [MinD](https://arxiv.org/abs/2506.18897) | MinD: Learning A Dual-System World Model for Real-Time Planning and Implicit Risk Analysis | arXiv 2025 | 6/2025 | `robot-only; exo; action-labeled; scripted` | [Project](https://manipulate-in-dream.github.io/) / [Code](https://github.com/manipulate-in-dream/MinD) | Bridge: `World Model / Planning / Risk`. Strict VAM: `Partial`. Dual-system world model for real-time planning and implicit failure prediction. |
| [iMoWM](https://arxiv.org/abs/2510.09036) | iMoWM: Taming Interactive Multi-Modal World Model for Robotic Manipulation | arXiv 2025 | 10/2025 | `robot-only; exo; action-labeled; scripted` | [Project](https://xingyoujun.github.io/imowm/) | Bridge: `World Model / Joint Video-Action`. Strict VAM: `Partial`. Interactive multi-modal world model where rollout quality is the central control object. |
| [VLAW](https://arxiv.org/abs/2602.12063) | VLAW: Iterative Co-Improvement of Vision-Language-Action Policy and World Model | arXiv 2026 | 2/2026 | `robot-only; exo; action-labeled; scripted` | [Project](https://sites.google.com/view/vlaw-arxiv) / [Code](https://github.com/Robert-gyj/Ctrl-World) / [Ckpt](https://huggingface.co/yjguo/Ctrl-World) | Bridge: `World Model / Policy-World Co-Training`. Strict VAM: `Partial`. Iterative co-improvement of policy and world model instead of one-way augmentation. |
| [World-VLA-Loop](https://arxiv.org/abs/2602.06508) | World-VLA-Loop: Closed-Loop Learning of Video World Model and VLA Policy | arXiv 2026 | 2/2026 | `robot-only; exo; action-labeled; scripted` | [Project](https://showlab.github.io/World-VLA-Loop/) / [Code](https://github.com/showlab/World-VLA-Loop) | Bridge: `World Model / Policy-World Co-Training`. Strict VAM: `Partial`. Closed-loop learning where the world model and policy continuously improve each other. |
| [Fast-WAM](https://arxiv.org/abs/2603.16666) | Fast-WAM: Do World Action Models Need Test-time Future Imagination? | arXiv 2026 | 3/2026 | `robot-only; exo; action-labeled; scripted` | [Project](https://yuantianyuan01.github.io/FastWAM/) / [Code](https://github.com/yuantianyuan01/FastWAM) | Bridge: `World Model / Joint Video-Action`. Strict VAM: `Partial`. Separates training-time video co-modeling from test-time control and shows fast action-only inference can retain most WAM gains. |
| [Privileged Foresight Distillation](https://arxiv.org/abs/2604.25859) | Privileged Foresight Distillation: Zero-Cost Future Correction for World Action Models | arXiv 2026 | 4/2026 | `robot-only; exo; action-labeled; scripted` | - | Bridge: `World Model / Privileged Future Correction`. Strict VAM: `Partial`. Support: `Yes`. Distills a teacher that sees ground-truth future observations into a current-only residual adapter, preserving Fast-WAM-style action-only deployment with zero test-time future-generation cost. |
| [GigaWorld-Policy](https://arxiv.org/abs/2603.17240) | GigaWorld-Policy: An Efficient Action-Centered World--Action Model | arXiv 2026 | 3/2026 | `robot-only; exo; action-labeled; scripted` | [Project](https://gigaai-research.github.io/GigaWorld-Policy/) / [Code](https://github.com/open-gigaai/giga-world-policy) | Bridge: `World Model / Action-Centered Joint Video-Action`. Strict VAM: `Partial`. Trains under coupled action-video supervision but makes explicit future-video generation optional at inference, clarifying the gap between WAM training and test-time control. |
| [World-Value-Action Model](https://arxiv.org/abs/2604.14732) | World-Value-Action Model: Implicit Planning for Vision-Language-Action Systems | arXiv 2026 | 4/2026 | `robot-only; exo; action-labeled; scripted` | - | Bridge: `World Model / Value-Guided Joint Video-Action`. Strict VAM: `Partial`. Couples imagined world trajectories, learned value filtering, and action decoding so future imagination is selected for controllability before becoming robot actions. |
| [STARRY](https://arxiv.org/abs/2604.26848) | STARRY: Spatial-Temporal Action-Centric World Modeling for Robotic Manipulation | arXiv 2026 | 4/2026 | `robot-only; multi-view; action-labeled; scripted` | - | Bridge: `World Model / Action-Centric Spatial-Temporal Foresight`. Strict VAM: `Partial`. Predicts future spatial-temporal latents and uses geometry-aware action attention to bind control to future-relevant scene tokens. |
| [Multi-View Video Diffusion Policy (MV-VDP)](https://arxiv.org/abs/2604.03181) | Multi-View Video Diffusion Policy: A 3D Spatio-Temporal-Aware Video Action Model | arXiv 2026 | 4/2026 | `robot-only; multi-view; action-labeled; scripted` | [Project](https://lpy1219.github.io/MV-VDP-Web/) | Bridge: `Explicit 2D / Multi-view Heatmap + Video`. Strict VAM: `Yes`. Jointly predicts multi-view RGB and heatmap videos before decoding executable action chunks. |
| [Action Images](https://arxiv.org/abs/2604.06168) | Action Images: End-to-End Policy Learning via Multiview Video Generation | arXiv 2026 | 4/2026 | `robot-only; multi-view; action-labeled; scripted` | [Project](https://actionimages.github.io/) / [Code](https://github.com/UMass-Embodied-AGI/ActionImages) | Bridge: `Explicit 2D / Pixel-Grounded Action Images`. Strict VAM: `Yes`. Converts 7-DoF actions into multi-view action videos so a video backbone can jointly model observations and executable action images. |
| [VAG](https://arxiv.org/abs/2604.09330) | VAG: Dual-Stream Video-Action Generation for Embodied Data Synthesis | arXiv 2026 | 4/2026 | `robot-only; exo; action-labeled; scripted` | - | Bridge: `World Model / Synthetic Video-Action Data`. Strict VAM: `Partial`. Support: `Yes`. Uses dual-stream video-action generation to synthesize paired embodied trajectories for downstream policy training. |
| [AIM](https://arxiv.org/abs/2604.11135) | AIM: Intent-Aware Unified world action Modeling with Spatial Value Maps | arXiv 2026 | 4/2026 | `robot-only; multi-view; action-labeled; scripted` | - | Bridge: `Explicit 2D / Spatial Value Map + Joint Video-Action`. Strict VAM: `Yes`. Optimization tag: `VAM-RL`. Jointly predicts future observations, action-based spatial value maps, and actions while forcing future information through the value-map interface before self-distilling the action head with RL. |
| [OA-WAM](https://arxiv.org/abs/2605.06481) | Object-Addressable World Action Model for Robust Robot Manipulation | arXiv 2026 | 5/2026 | `robot-only; exo; action-labeled; scripted` | - | Bridge: `World Model / Object-Addressable Joint Video-Action`. Strict VAM: `Partial`. Splits object slots into fixed identity addresses and time-varying content so action decoding can query target objects robustly. |
| [HarmoWAM](https://arxiv.org/abs/2605.10942) | HarmoWAM: Harmonizing Generalizable and Precise Manipulation via Adaptive World Action Models | arXiv 2026 | 5/2026 | `robot-only; exo; action-labeled; scripted` | [Project](https://elbb-yu.github.io/HarmoWAM/) / [Code](https://github.com/xuanxuanzzzii/HarmoWAM) | Bridge: `World Model / Adaptive Expert Gating`. Strict VAM: `Partial`. Combines reactive/exploratory and predictive/precise WAM experts with process-adaptive phase gating for robust manipulation. |
| [NoiseGate](https://arxiv.org/abs/2605.07794) | NoiseGate: Learning Per-Latent Timestep Schedules as Information Gating in World Action Models | arXiv 2026 | 5/2026 | `robot-only; exo; action-labeled; scripted` | - | Bridge: `World Model / Per-Latent Denoising Gate`. Strict VAM: `Partial`. Optimization tag: `VAM-RL`. Learns a GRPO-trained per-future-latent timestep schedule so action tokens read imagined futures according to task-shaped reliability. |
| [Point Tracking Improves WAM](https://arxiv.org/abs/2605.23856) | Point Tracking Improves World Action Models | arXiv 2026 | 5/2026 | `robot-only; exo; action-labeled; scripted` | - | Bridge: `Explicit 2D / Point-Track Future State`. Strict VAM: `Partial`. Adds point tracks and visibility to joint video-action denoising so action tokens can condition on persistent object motion rather than appearance-only future latents. |
| [SANTS](https://arxiv.org/abs/2605.27947) | SANTS: A State-Adaptive Scheduler for World Action Models | arXiv 2026 | 5/2026 | `robot-only; exo; action-labeled; scripted` | [Project](https://advanced-robotics-lab.github.io/SANTS/) / [Code](https://github.com/Advanced-Robotics-Lab/SANTS-code) | Bridge: `World Model / State-Adaptive Denoising`. Strict VAM: `No`. Support: `Yes`. Optimization tag: `VAM-RL`. Learns when to stop or jump along the video denoising trajectory according to downstream action quality rather than visual fidelity alone. |
| [WALL-WM](https://arxiv.org/abs/2606.01955) | WALL-WM: Carving World Action Modeling at the Event Joints | arXiv 2026 | 6/2026 | `robot-only; multi-view; action-labeled; scripted` | [Code](https://github.com/X-Square-Robot/wall-x) | Bridge: `World Model / Event-Centric Joint Video-Action`. Strict VAM: `Partial`. Aligns video, language, and action around embodied event joints, making the event boundary the unit of joint world-action modeling. |
| [Flash-WAM](https://arxiv.org/abs/2606.05254) | Flash-WAM: Modality-Aware Distillation for World Action Models | arXiv 2026 | 6/2026 | `robot-only; exo; action-labeled; scripted` | [Project](https://flashwam.github.io/) / [Code](https://github.com/NU-World-Model-Embodied-AI/Flash-WAM) | Bridge: `World Model / Distilled Joint Video-Action`. Strict VAM: `No`. Support: `Yes`. Distillation and acceleration paper that directly supports deployable WAM inference rather than introducing a new strict VAM backbone. |
| [AdaWAM](https://arxiv.org/abs/2606.07089) | Dreaming when Necessary: Advancing World Action Models with Adaptive Multi-Modal Reasoning | arXiv 2026 | 6/2026 | `robot-only; exo; action-labeled; scripted` | [Project](https://adawam.github.io/) | Bridge: `World Model / Adaptive Future Imagination`. Strict VAM: `Partial`. Routes each decision among language reasoning, visual dreaming, and direct action, so future imagination is used only when it is action-useful. |
| [FAWAM](https://arxiv.org/abs/2606.08555) | FAWAM: Force-Aware World Action Models for Closed-Loop Contact-Rich Manipulation | arXiv 2026 | 6/2026 | `robot-only; visuotactile; action-labeled; scripted` | - | Bridge: `Physics / Force-Aware Joint World-Action`. Strict VAM: `Partial`. Extends WAM-style future/action modeling with contact force prediction and closed-loop force discrepancy correction. |
| [AHA-WAM](https://arxiv.org/abs/2606.09811) | AHA-WAM: Asynchronous Horizon-Adaptive World-Action Modeling with Observation-Guided Context Routing | arXiv 2026 | 6/2026 | `robot-only; exo; action-labeled; scripted` | [Project](https://serene-sivy.github.io/aha-wam/) | Bridge: `World Model / Asynchronous Horizon-Adaptive WAM`. Strict VAM: `Partial`. Decouples world-prediction horizon from action-execution horizon and routes context according to observation needs. |
| [Efficient-WAM](https://arxiv.org/abs/2606.10040) | Efficient-WAM: A 1B-Parameter World-Action Model with Low-Cost Future Imagination | arXiv 2026 | 6/2026 | `robot-only; exo; action-labeled; scripted` | - | Bridge: `World Model / Low-Cost Future Imagination`. Strict VAM: `Partial`. Compresses model size, future-token budget, and denoising cost for low-latency WAM-style control. |
| [HiMem-WAM](https://arxiv.org/abs/2606.10363) | HiMem-WAM: Hierarchical Memory-Gated World Action Models for Robotic Manipulation | arXiv 2026 | 6/2026 | `robot-only; exo; action-labeled; scripted` | - | Bridge: `World Model / Hierarchical Memory-Gated WAM`. Strict VAM: `Partial`. Combines memory, skill-level latent structure, latent actions, and future modeling through boundary-triggered gates. |
| [EWAM](https://arxiv.org/abs/2606.12690) | EWAM: An Enhanced World Action Model for Closed-Loop Online Adaptation in Embodied Intelligence | arXiv 2026 | 6/2026 | `robot-only; exo; action-labeled; scripted` | - | Bridge: `Execution / Closed-Loop Online WAM Adaptation`. Strict VAM: `No`. Support: `Yes`. Wraps a WAM with memory, anomaly handling, rollback, filtering, and online adaptation; useful, but more support architecture than core VAM definition. |
| [MaskWAM](https://arxiv.org/abs/2606.13515) | MaskWAM: Unifying Mask Prompting and Prediction for World-Action Models | arXiv 2026 | 6/2026 | `robot-only; exo; action-labeled; scripted` | [Project](https://hanyangyu1021.github.io/maskwam.github.io/) / [Code](https://github.com/hanyangyu1021/maskwam) | Bridge: `Explicit 2D / Mask-Prompted Joint World-Action`. Strict VAM: `Partial`. Uses first-frame masks as prompts and future masks as prediction targets, giving joint WAM training a clean structured 2D bridge. |
| [Making Foresight Actionable](https://arxiv.org/abs/2606.12217) | Making Foresight Actionable: Repurposing Representation Alignment in World Action Models | arXiv 2026 | 6/2026 | `robot-only; exo; action-labeled; scripted` | - | Bridge: `World Model / Action-Grounded Foresight Alignment`. Strict VAM: `No`. Support: `Yes`. Diagnoses the gap between plausible future latents and action-readable foresight, then aligns representations so WAM imagination can actually steer the action head. |
| [World Pilot](https://arxiv.org/abs/2606.12403) | World Pilot: Steering Vision-Language-Action Models with World-Action Priors | arXiv 2026 | 6/2026 | `robot-only; exo; action-labeled; scripted` | - | Bridge: `World Model / World-Action Prior Injection`. Strict VAM: `Partial`. Injects frozen WAM future latents and coarse action priors into a VLA, making it a WAM-to-VLA steering bridge. |
| [C3ache](https://arxiv.org/abs/2606.08962) | C3ache: Accelerating World Action Models with Cross Inference Chunk Cache | arXiv 2026 | 6/2026 | `robot-only; exo; action-labeled; scripted` | - | Bridge: `World Model / Cross-Chunk Action Cache`. Strict VAM: `No`. Support: `Yes`. Inference systems paper that caches cross-chunk WAM computation to accelerate the action expert without changing the core modeling objective. |
| [Donk](https://arxiv.org/abs/2606.03868) | Unified Video-Action Joint Denoising for Dexterous Action and Data Generation | arXiv 2026 | 6/2026 | `dexterous-hand; exo; action-labeled; scripted` | - | Bridge: `World Model / Joint Video-Action`. Strict VAM: `Yes`. Jointly denoises video tokens and MANO hand-action tokens, serving both observation-conditioned dexterous action prediction and text-conditioned video-action data generation. |
| [Dream-Tac](https://arxiv.org/abs/2606.08737) | Dream-Tac: A Unified Tactile World Action Model for Contact-Rich Robot Manipulation | arXiv 2026 | 6/2026 | `robot-only; visuotactile; action-labeled; scripted` | [Code](https://github.com/LYFCLOUDFAN/Dream-Tac) | Bridge: `Physics / Visuotactile Joint World-Action`. Strict VAM: `Partial`. Extends joint WAM prediction to future visual observations, tactile dynamics, and robot actions with contact-aware fusion. |
| [RepWAM](https://arxiv.org/abs/2606.13674) | RepWAM: World Action Modeling with Representation Visual-Action Tokenizers | arXiv 2026 | 6/2026 | `robot-only; exo; action-labeled; scripted` | [Project](https://wdrink.github.io/RepWAM/) / [Code](https://github.com/wdrink/RepWAM) | Bridge: `World Model / Representation Visual-Action Tokens`. Strict VAM: `Partial`. Learns semantic visual-action latents, then causally models future visual states and latent actions before robot-action adaptation. |
| [MV-WAM](https://arxiv.org/abs/2606.21088) | MV-WAM: Manifold-Aware World Action Model with Value Augmentation | arXiv 2026 | 6/2026 | `robot-only; exo; action-labeled; scripted` | - | Bridge: `World Model / Value-Augmented Joint Video-Action`. Strict VAM: `Partial`. Jointly models visual prediction, action generation, and value estimation with manifold-aware objectives and value-guided recovery. |
| [Tactile-WAM](https://arxiv.org/abs/2606.26663) | Tactile-WAM: Touch-Aware World Action Model with Tactile Asymmetric Attention | arXiv 2026 | 6/2026 | `robot-only; visuotactile; action-labeled; scripted` | - | Bridge: `Physics / Touch-Aware Joint World-Action`. Strict VAM: `Partial`. Uses asymmetric tactile attention so action tokens can use contact cues while visual prediction is protected from tactile-token pollution. |
| [DIM-WAM](https://arxiv.org/abs/2606.27677) | DIM-WAM: World-Action Modeling with Diverse Historical Event Memory | arXiv 2026 | 6/2026 | `robot-only; exo; action-labeled; scripted` | [Project](https://wangkai-casia.github.io/dim-wam/) / [Code](https://github.com/wangkai-casia/dim-wam) | Bridge: `World Model / Diverse Historical Event Memory`. Strict VAM: `Partial`. Conditions video and action denoising on multi-scale event memories and progress supervision for long-horizon manipulation. |
| [A2World](https://arxiv.org/abs/2606.29501) | Learning Transferable Dynamics Priors from Action to World Modeling | ECCV 2026 | 6/2026 | `robot-only; multi-view; action-labeled; scripted` | [Code](https://github.com/LogosRoboticsGroup/A2World) | Bridge: `World Model / Action-to-Video Prior + Joint Video-Action`. Strict VAM: `Partial`. Pretrains a multi-view action-conditioned video world model on large robot trajectories, then adapts the same dynamics prior into a simulator and a language-conditioned joint video-action policy. |
| [Bridge-WA](https://arxiv.org/abs/2607.02195) | Bridge-WA: Predicting Where and How the World Changes for Robotic Action | arXiv 2026 | 7/2026 | `robot-only; exo; action-labeled; structured future-change maps` | [Project](https://hcplab-sysu.github.io/BRIDGE-WA/) / [Code](https://github.com/HCPLab-SYSU/BRIDGE-WA) / [Ckpt](https://huggingface.co/baiyu858/Bridge-WA-VLABench) | Bridge: `Explicit 2D / Future Tokens + Change Maps + Motion-Flow`. Strict VAM: `Partial`. Distills a frozen world teacher into future tokens, change maps, and motion-flow maps that condition an action transformer while removing the teacher at inference. |
| [VT-WAM](https://arxiv.org/abs/2607.02503) | VT-WAM: Visual-Tactile World Action Model for Contact-Rich Manipulation | arXiv 2026 | 7/2026 | `robot-only; visuotactile; action-labeled; scripted` | - | Bridge: `World Model / Joint Visual-Tactile-Action`. Strict VAM: `Yes`. Jointly predicts future visual observations, tactile deformation, and actions in one flow-matching objective, with contact-aware tactile evidence feeding the action queries. |
| [Qantara](https://arxiv.org/abs/2607.04978) | Qantara: Bridge-Flow Training for Multi-Paradigm JEPA Control | arXiv 2026 | 7/2026 | `robot-only; exo; action-labeled; scripted` | [Project](https://corl-team.github.io/qantara/) | Bridge: `World Model / Shared JEPA State-Action Flow`. Strict VAM: `Partial`. Couples a Brownian bridge over state latents with action flow matching so one checkpoint supports latent planning, behavior-cloning action sampling, and video-inverse action recovery. |
| [LingBot-VA 2.0](https://arxiv.org/abs/2607.08639) | Native Video-Action Pretraining for Generalizable Robot Control | arXiv 2026 | 7/2026 | `co-training; exo; action-free + action-labeled; scripted` | - | Bridge: `World Model / Joint Causal Video-Action Latent`. Strict VAM: `Yes`. Performs native causal video-action pretraining with a semantic visual-action tokenizer, then generates future latents and executable actions asynchronously while re-grounding on new observations. |
| [FlowWAM](https://arxiv.org/abs/2607.13017) | FlowWAM: Optical Flow as a Unified Action Representation for World Action Models | arXiv 2026 | 7/2026 | `co-training; exo; action-free + action-labeled; scripted` | [Project](https://flow-wam.github.io/) / [Code](https://github.com/YixiangChen515/FlowWAM) | Bridge: `Explicit 2D / Optical-Flow Video Action`. Strict VAM: `Yes`. Uses optical-flow video as a video-native action representation inside a shared generator, unifying policy mode, action-conditioned future prediction, and action-free video pretraining. |

<a id="route-3-unified-video-action-world-action-models"></a>
# 3️⃣ Route 3: Unified Video-Action / World-Action Models

3️⃣ `R3` is where unified training and action-only deployment become explicit design targets.

| Method | Title | Venue | Date | Data tags | Code/Project | Key feature/finding |
| --- | --- | --- | --- | --- | --- | --- |
| [UWM](https://arxiv.org/abs/2504.02792) | Unified World Models: Coupling Video and Action Diffusion for Pretraining on Large Robotic Datasets | RSS 2025 | 4/2025 | `robot-only; exo; action-free + action-labeled; scripted` | [Project](https://weirdlabuw.github.io/uwm/) / [Code](https://github.com/WEIRDLabUW/unified-world-model) / [Ckpt](https://drive.google.com/drive/folders/1M4AuVLMRpSwOf_YAp56bV9AqyZI9ul6g?usp=sharing) | Bridge: `World Model / Mixed Action-Labeled + Action-Free Training`. Strict VAM: `Yes`. Unified formulation with explicit support for mixed training on action-labeled trajectories and action-free videos. |
| [UVA](https://arxiv.org/abs/2503.00200) | Unified Video Action Model | RSS 2025 | 3/2025 | `robot-only; exo; action-free + action-labeled; scripted` | [Project](https://unified-video-action-model.github.io/) / [Code](https://github.com/ShuangLI59/unified_video_action) / [Ckpt](https://drive.google.com/file/d/11c2VrmaRp48yw__5A5xpcu8EPzkexHSi/view?usp=sharing) | Bridge: `World Model / Shared Latent`. Strict VAM: `Partial`. Joint training with decoupled action-only inference. |
| [UniVLA](https://arxiv.org/abs/2505.06111) | UniVLA: Learning to Act Anywhere with Task-centric Latent Actions | RSS 2025 | 5/2025 | `co-training; exo; action-free + action-labeled; scripted` | [Code](https://github.com/OpenDriveLab/UniVLA) | Bridge: `Latent Action`. Strict VAM: `No`. Boundary case: learns task-centric latent actions from web-scale cross-embodiment videos and decodes them efficiently into executable control. |
| [Motus](https://arxiv.org/abs/2512.13030) | Motus: A Unified Latent Action World Model | arXiv 2025 | 12/2025 | `co-training; exo; action-free + action-labeled; scripted` | [Project](https://motus-robotics.github.io/motus) / [Code](https://github.com/thu-ml/Motus) / [Ckpt](https://huggingface.co/motus-robotics/Motus) | Bridge: `Latent Action`. Strict VAM: `Yes`. One of the clearest latent-action unified world-model papers. |
| [LAWM](https://arxiv.org/abs/2601.05230) | Learning Latent Action World Models In The Wild | arXiv 2026 | 1/2026 | `co-training; exo; action-free + action-labeled; unscripted` | - | Bridge: `Latent Action`. Strict VAM: `Partial`. Boundary case: `R3`-adjacent latent-action world-model paper that stress-tests embodiment invariance on wild video. |
| [JALA](https://arxiv.org/abs/2602.21736) | Joint-Aligned Latent Action: Towards Scalable VLA Pretraining in the Wild | CVPR 2026 | 2/2026 | `co-training; ego; action-free + action-labeled; unscripted` | [Code](https://github.com/BeingBeyond/JALA) | Bridge: `Latent Action`. Strict VAM: `No`. Boundary case: jointly aligns predictive embeddings, inverse-dynamics latents, and real actions to turn in-the-wild human video into scalable action-centric supervision. |
| [EC-Flow](https://arxiv.org/abs/2507.06224) | EC-Flow: Enabling Versatile Robotic Manipulation from Action-Unlabeled Videos via Embodiment-Centric Flow | ICCV 2025 | 7/2025 | `co-training; exo; action-free + action-labeled; scripted` | [Project](https://ec-flow1.github.io/) / [Code](https://github.com/YixiangChen515/EC-Flow) / [Ckpt](https://huggingface.co/YixiangChen/EC-Flow) | Bridge: `Latent Action / Embodiment-Centric Flow`. Strict VAM: `No`. Boundary case: bridge-heavy method that turns action-unlabeled videos into executable robot supervision through embodiment-centric motion flow. |
| [Learning Skills from Action-Free Videos](https://arxiv.org/abs/2512.20052) | Learning Skills from Action-Free Videos | arXiv 2025 | 12/2025 | `co-training; exo; action-free + action-labeled; scripted` | - | Bridge: `Latent Action / Skills`. Strict VAM: `No`. Boundary case: skill-abstraction route from action-free visual interaction data, better read as `R3`-adjacent than as a canonical unified anchor. |
| [Being-H0.7](https://research.beingbeyond.com/being-h07) | Being-H0.7: A Latent World-Action Model from Egocentric Videos | BeingBeyond 2026 | 4/2026 | `co-training; ego; action-free + action-labeled; unscripted` | [Project](https://research.beingbeyond.com/being-h07) / [Code](https://github.com/BeingBeyond/Being-H) | Bridge: `World Model / Future-Aware Latent Query`. Strict VAM: `Partial`. Distills a future-aware posterior branch into deployable prior latent queries, keeping action-only inference while using egocentric video to shape latent world-action reasoning. |
| [X-WAM](https://arxiv.org/abs/2604.26694) | X-WAM: Unified 4D World Action Modeling from Video Priors with Asynchronous Denoising | arXiv 2026 | 4/2026 | `robot-only; multi-view; action-labeled; scripted` | [Project](https://sharinka0715.github.io/X-WAM/) | Bridge: `Explicit 3D / Unified 4D World-Action`. Strict VAM: `Yes`. Unifies multi-view RGB-D future generation, 4D reconstruction, and action chunk prediction, using asynchronous denoising to keep world modeling and low-latency control in one WAM. |
| [MotuBrain](https://arxiv.org/abs/2604.27792) | MotuBrain: An Advanced World Action Model for Robot Control | arXiv 2026 | 4/2026 | `co-training; multi-view; action-free + action-labeled; scripted` | [Project](https://www.shengshu.com/en/motubrain) | Bridge: `World Model / Unified WAM Fast Path`. Strict VAM: `Yes`. Unifies policy learning, world modeling, video generation, IDM, and joint video-action prediction while using V2A/action-only suffix inference, caching, quantization, and asynchronous chunks for deployable latency. |
| [CKT-WAM](https://arxiv.org/abs/2605.06247) | CKT-WAM: Parameter-Efficient Context Knowledge Transfer Between World Action Models | arXiv 2026 | 5/2026 | `robot-only; exo; action-labeled; scripted` | [Code](https://github.com/YuhuaJiang2002/CKT-WAM) | Bridge: `World Model / Teacher Context Transfer`. Strict VAM: `Partial`. Support: `Yes`. Compresses a heavy teacher WAM's intermediate world-action states into text-side context tokens for a lighter student WAM. |
| [ALAM](https://arxiv.org/abs/2605.10819) | ALAM: Algebraically Consistent Latent Transitions for Vision-Language-Action Models | arXiv 2026 | 5/2026 | `co-training; exo; action-free + action-labeled; scripted` | - | Bridge: `Latent Action / Algebraic Transition`. Strict VAM: `No`. Boundary case: learns additive and reversible latent transitions and co-generates them with executable actions inside a joint flow policy. |
| [DiLA](https://arxiv.org/abs/2605.15725) | DiLA: Disentangled Latent Action World Models | ICML 2026 | 5/2026 | `co-training; exo; action-free + action-labeled; scripted` | [Project](http://disentangled-latent-action-world-models.github.io) | Bridge: `Latent Action / Disentangled World Model`. Strict VAM: `Partial`. Learns a content-structure disentangled latent-action world model where the action bottleneck drives structure dynamics while a separate content memory preserves visual detail for rollout, transfer, and planning. |
| [UWM-JEPA](https://arxiv.org/abs/2605.25313) | UWM-JEPA: Predictive World Models That Imagine in Belief Space | arXiv 2026 | 5/2026 | `world-model; action-conditioned; latent-belief` | [Code](https://github.com/santoshkumarradha/uwm-jepa) | Bridge: `World Model / Belief-Space Latent Imagination`. Strict VAM: `No`. Support: `Yes`. Uses density-matrix belief latents and unitary predictors to preserve hidden continuation information during JEPA-style action-conditioned blind rollouts. |
| [tau0-WM](https://arxiv.org/abs/2606.01027) | tau0-WM: A Unified Video-Action World Model for Robotic Manipulation | arXiv 2026 | 6/2026 | `robot-only; exo; action-labeled; scripted` | [Project](https://finch.agibot.com/research/tau0-wm) / [Code](https://github.com/sii-research/tau-0-wm) | Bridge: `World Model / Unified Video-Action`. Strict VAM: `Partial`. R3/R2 boundary: unifies future video prediction, action candidates, progress checking, and test-time correction in one video-action world model. |
| [CLAW](https://arxiv.org/abs/2606.04130) | CLAW: Learning Continuous Latent Action World Models via Adversarial Latent Regularization | arXiv 2026 | 6/2026 | `co-training; exo; action-free + action-labeled; scripted` | - | Bridge: `Latent Action / Continuous Action-Free World Model`. Strict VAM: `Partial`. Learns continuous latent actions from action-free videos while regularizing them not to encode the future image directly. |
| [World-Language-Action Model](https://arxiv.org/abs/2606.05979) | World-Language-Action Model for Unified World Modeling, Language Reasoning, and Action Synthesis | arXiv 2026 | 6/2026 | `robot-only; exo; action-labeled; scripted` | [Code](https://github.com/SJTU-DENG-Lab/WLA) | Bridge: `World Model / Unified World-Language-Action`. Strict VAM: `Partial`. Unified route paper tying world modeling, language-state reasoning, and action synthesis while preserving an action-oriented inference path. |
| [LARA](https://arxiv.org/abs/2606.07100) | LARA: Latent Action Representation Alignment for Vision-Language-Action Models | arXiv 2026 | 6/2026 | `co-training; exo; action-free + action-labeled; scripted` | - | Bridge: `Latent Action / VLA Hidden-State Alignment`. Strict VAM: `No`. Boundary/support paper that aligns latent-action representations with VLA hidden states rather than proposing a full world-action model. |
| [WAM4D](https://arxiv.org/abs/2606.14048) | WAM4D: Fast 4D World Action Model via Spatial Register Tokens | arXiv 2026 | 6/2026 | `robot-only; explicit-4D; action-labeled; scripted` | - | Bridge: `Explicit 3D / Spatial Register Tokens`. Strict VAM: `Partial`. R3/R2 boundary: transfers geometric priors through training-time future-depth register readouts, then removes the register branch for lightweight action inference. |
| [MemoryWAM](https://arxiv.org/abs/2606.20562) | MemoryWAM: Efficient World Action Modeling with Persistent Memory | arXiv 2026 | 6/2026 | `robot-only; exo; action-labeled; scripted` | [Project](https://yangsizhe.github.io/MemoryWAM/) / [Code](https://github.com/yangsizhe/MemoryWAM) | Bridge: `World Model / Persistent Memory`. Strict VAM: `Partial`. Uses hybrid recent-frame, event-anchor, and gist-token memory so video generation is bypassed at inference while action decoding retains long-range context. |
| [ABot-M0.5](https://arxiv.org/abs/2607.00678) | ABot-M0.5: Unified Mobility-and-Manipulation World Action Model | arXiv 2026 | 7/2026 | `robot-only; mobile-manipulation; multi-view; action-labeled` | [Code](https://github.com/amap-cvlab/ABot-Manipulation) | Bridge: `World Model / Mobility-Manipulation WAM`. Strict VAM: `Yes`. Aligns future video latents, frame-level latent actions, and heterogeneous mobility/manipulation action spaces with Dual-level MoT and Dream Forcing for action-only deployment. |
| [WSA1](https://arxiv.org/abs/2607.03941) | WSA1: A 3D-Centric World-Spatial-Action Model for Generalizable Robot Control | arXiv 2026 | 7/2026 | `co-training; multi-view + ego; action-free + action-labeled; scripted` | [Code](https://github.com/zaleni/WSA) | Bridge: `Explicit 3D / World-Spatial-Action`. Strict VAM: `Partial`. Jointly learns action-conditioned 3D world prediction, 3D-grounded visual thinking, and 3D inverse dynamics so future scene tokens and action tokens constrain each other. |
| [HALO-WA](https://arxiv.org/abs/2607.04265) | HALO-WA: Hybrid-Attention Latent-Guided Online Reinforcement Learning for World-Action Models | arXiv 2026 | 7/2026 | `robot-only; multi-view; action-labeled + online RL; scripted` | [Code](https://github.com/YeanRoot/HALO-WA) | Bridge: `Optimization / Latent-Guided Online WAM RL`. Strict VAM: `No`. Support: `Yes`. Optimization tag: `VAM-RL`. Freezes a WA backbone and trains a lightweight actor-critic adapter over reference actions and visual latents for real-world precision adaptation. |
| [DSWAM](https://arxiv.org/abs/2607.04927) | DSWAM: A Dual-System World Action Foundation Model for Fine-Grained Robot Manipulation | arXiv 2026 | 7/2026 | `robot-only; multi-view; dual-arm; action-labeled` | - | Bridge: `World Model / Dual-System Action-Only WAM`. Strict VAM: `Partial`. Uses a default WAM executor with training-time video co-modeling and action-only inference, plus an optional VLM subtask planner and real-time deployment path. |
| [InternVLA-A1.5](https://arxiv.org/abs/2607.04988) | InternVLA-A1.5: Unifying Understanding, Latent Foresight, and Action for Compositional Generalization | arXiv 2026 | 7/2026 | `co-training; exo; action-labeled + multimodal; scripted` | [Project](https://internrobotics.github.io/internvla-a15.github.io) / [Code](https://github.com/InternRobotics/InternVLA-A-series) / [Ckpt](https://huggingface.co/collections/InternRobotics/internvla-a15) | Bridge: `World Model / Training-Time Latent Foresight`. Strict VAM: `Partial`. Uses learnable foresight tokens supervised by a frozen video teacher, then discards the video branch at inference and deploys a flow-matching action expert. |
| [WorldBagel](https://arxiv.org/abs/2607.03461) | WorldBagel: Uncovering the Power of Unified Multimodal Models for Vision-Language-Action-World Modeling | arXiv 2026 | 7/2026 | `co-training; exo; action-free + action-labeled; scripted` | - | Bridge: `World Model / Unified Vision-Language-Action-World`. Strict VAM: `Partial`. Unifies vision, language, action, and world representations in one multimodal framework and validates the shared formulation across manipulation benchmarks and real robots. |
| [MECo-WAM](https://arxiv.org/abs/2607.05468) | Learning 4D Geometric Priors for Inference-Efficient World Action Models | arXiv 2026 | 7/2026 | `robot-only; exo; action-labeled; scripted` | - | Bridge: `Explicit 3D/4D / Training-Time Geometric Prior`. Strict VAM: `Partial`. Co-trains video, action, and 4D geometry experts, distills temporal geometric structure into the video-action path, and removes the auxiliary 4D expert at deployment. |
| [EgoWAM](https://arxiv.org/abs/2607.08436) | EgoWAM: World Action Models Beyond Pixels with In-the-Wild Egocentric Human Data | arXiv 2026 | 7/2026 | `co-training; ego; action-free + action-labeled; unscripted` | [Project](https://gatech-rl2.github.io/egowam.github.io/) | Bridge: `World Model / Human-Robot Predictive Co-Training`. Strict VAM: `Partial`. Co-trains robot actions with pixel, semantic-feature, and 3D-flow world targets while using in-the-wild egocentric human video to improve dynamics transfer. |
| [Lumo-2](https://arxiv.org/abs/2607.11270) | Towards Predictive, Aligned, and Scalable Robot Learning | arXiv 2026 | 7/2026 | `robot-only; exo; action-labeled; scripted` | - | Bridge: `World Model / Latent Physical Dynamics`. Strict VAM: `Yes`. Generates actions by reasoning over physically grounded latent world dynamics and progressively aligns action representations with world, vision, and language geometry. |
| [WALA](https://arxiv.org/abs/2607.11397) | WALA: Learning Executable Latent Actions from Action-Labeled Demonstrations and Action-Free Videos | arXiv 2026 | 7/2026 | `co-training; exo; action-free + action-labeled; scripted` | [Project](https://liujiahao2077.github.io/WALA.github.io/) / [Code](https://github.com/liujiahao2077/WALA) | Bridge: `Latent Action / Semantic-Geometric Future Delta`. Strict VAM: `Partial`. Learns semantic-geometric latent actions from action-free video and jointly supervises executable robot actions, latent targets, and future DINO/depth dynamics during policy training. |
| [GigaWorld-Policy-0.5](https://arxiv.org/abs/2607.13960) | GigaWorld-Policy-0.5: A Faster and Stronger WAM Empowered by AutoResearch | arXiv 2026 | 7/2026 | `robot-only; exo; action-labeled; scripted` | [Project](https://open-gigaai.github.io/giga-world-policy/) / [Code](https://github.com/open-gigaai/giga-world-policy) | Bridge: `World Model / Mixed AC-WM + WAM Training`. Strict VAM: `Partial`. Mixes action-conditioned world modeling with world-action modeling during training, separates visual-dynamics and action experts, and deploys a fast action-only path. |
| [WorldScape Policy 2.0](https://arxiv.org/abs/2607.18840) | WorldScape Policy 2.0: Empowering Steerable World Action Modeling with Reasoning-Augmented Memory | arXiv 2026 | 7/2026 | `robot-only; exo; action-labeled + multimodal; scripted` | - | Bridge: `World Model / Joint Visual-State Action + Event Memory`. Strict VAM: `Partial`. Couples visual-state transitions and actions with causal short-term memory, event memory, and planning tokens for steerable long-horizon robotic manipulation. |

<a id="route-4-physical-intelligence-bridge-heavy-signals-and-evaluation"></a>
# 4️⃣ Route 4: Physical Intelligence, Bridge-Heavy Signals, and Evaluation

4️⃣ `R4` is a thematic route for physics, utility, safety, and functional realism rather than a separate backbone family. This section also includes bridge-heavy physical-intelligence papers whose primary route remains in R1–R3 but are grouped here for thematic coherence. Generic VLA papers without a concrete video/world-model or bridge-signal role are intentionally excluded.

| Method | Title | Venue | Date | Data tags | Code/Project | Key feature/finding |
| --- | --- | --- | --- | --- | --- | --- |
| [EWMBench](https://arxiv.org/abs/2505.09694) | EWMBench: Evaluating Scene, Motion, and Semantic Quality in Embodied World Models | arXiv 2025 | 5/2025 | `robot-only; exo; action-labeled; scripted` | [Code](https://github.com/AgibotTech/EWMBench) | Bridge: `Evaluation / Scene-Motion-Semantics`. Strict VAM: `No`. Benchmark that decomposes embodied world model quality into scene consistency, motion correctness, and semantic alignment instead of treating visual fidelity as enough. |
| [Genie Envisioner](https://arxiv.org/abs/2508.05635) | Genie Envisioner: A Unified World Foundation Platform for Robotic Manipulation | arXiv 2025 | 8/2025 | `robot-only; exo; action-labeled; scripted` | [Project](https://genie-envisioner.github.io/) / [Code](https://github.com/AgibotTech/Genie-Envisioner) | Bridge: `World Model / Policy + Simulator + Benchmark`. Strict VAM: `Partial`. Platform-style system combining GE-Act action decoding, GE-Sim action-conditioned simulation, and benchmark-oriented evaluation inside one video-generative stack. |
| [BridgeV2W](https://arxiv.org/abs/2602.03793) | BridgeV2W: Bridging Video Generation Models to Embodied World Models via Embodiment Masks | arXiv 2026 | 2/2026 | `robot-only; exo; action-labeled; scripted` | [Project](https://bridgev2w.github.io/) | Bridge: `Explicit 2D / Embodiment Mask + World Model`. Strict VAM: `Partial`. Converts video generation into embodied world modeling through embodiment-aware masks. |
| [MVISTA-4D](https://arxiv.org/abs/2602.09878) | MVISTA-4D: View-Consistent 4D World Model with Test-Time Action Inference for Robotic Manipulation | arXiv 2026 | 2/2026 | `robot-only; multi-view; action-labeled; scripted` | - | Bridge: `Explicit 3D / 4D World Model`. Strict VAM: `Partial`. View-consistent 4D world modeling with test-time action inference. |
| [WorldArena](https://arxiv.org/abs/2602.08971) | WorldArena: A Unified Benchmark for Evaluating Perception and Functional Utility of Embodied World Models | arXiv 2026 | 2/2026 | `robot-only; exo; action-labeled; scripted` | [Project](https://world-arena.ai/) / [Code](https://github.com/tsinghua-fib-lab/WorldArena) | Bridge: `Evaluation / Functional Utility`. Strict VAM: `No`. Benchmark for the gap between good-looking rollouts and useful control. |
| [WorldArena 2.0](https://arxiv.org/abs/2605.17912) | WorldArena 2.0: Extending Embodied World Model Benchmarking on Modality, Functionality and Platform | arXiv 2026 | 5/2026 | `robot-only; visuotactile; action-labeled; scripted` | [Project](https://world-arena.ai/) / [Code](https://github.com/tsinghua-fib-lab/WorldArena) | Bridge: `Evaluation / Multimodal Interactive Utility`. Strict VAM: `No`. Extends embodied world-model evaluation from vision-only offline utility to visuotactile prediction, online RL-environment use, and simulation-to-real platform transfer. |
| [ContactGaussian-WM](https://arxiv.org/abs/2602.11021) | ContactGaussian-WM: Learning Physics-Grounded World Model from Videos | arXiv 2026 | 2/2026 | `robot-only; multi-view; action-labeled; scripted` | [Project](https://contactgaussian-wm.github.io/) | Bridge: `Physics-Grounded World Model / Contact`. Strict VAM: `No`. Physics-grounded world model for contact-rich interaction. |
| [Geometry-aware 4D Video Generation](https://arxiv.org/abs/2507.01099) | Geometry-aware 4D Video Generation for Robot Manipulation | ICLR 2026 | 7/2025 | `robot-only; multi-view; action-labeled; scripted` | [Project](https://robot4dgen.github.io/) / [Code](https://github.com/lzylucy/4dgen) / [Ckpt](https://real.stanford.edu/4dgen/) | Bridge: `Explicit 3D / Multi-view 4D Geometry`. Strict VAM: `Partial`. Generates geometrically consistent multi-view RGB-D futures and recovers executable end-effector trajectories through off-the-shelf pose tracking. |
| [Inference-time Physics Alignment](https://arxiv.org/abs/2601.10553) | Inference-time Physics Alignment of Video Generative Models with Latent World Models | arXiv 2026 | 1/2026 | `robot-only; exo; action-labeled; scripted` | [Code](https://github.com/facebookresearch/WMReward) | Bridge: `Physics Alignment / World Model`. Strict VAM: `No`. Uses latent world models to align video generators at inference time. |
| [Ctrl-World](https://arxiv.org/abs/2510.10125) | Ctrl-World: A Controllable Generative World Model for Robot Manipulation | ICLR 2026 | 10/2025 | `robot-only; multi-view; action-labeled; scripted` | [Project](https://ctrl-world.github.io/) / [Code](https://github.com/Robert-gyj/Ctrl-World) | Bridge: `World Model / Controllable Multi-view Rollout`. Strict VAM: `Partial`. Support: `Yes`. Builds a controllable multi-view world model for policy evaluation and synthetic trajectory generation with pose-conditioned memory retrieval and frame-level action conditioning. |
| [WoVR](https://arxiv.org/abs/2602.13977) | WoVR: World Models as Reliable Simulators for Post-Training VLA Policies with RL | arXiv 2026 | 2/2026 | `robot-only; exo; action-labeled; scripted` | - | Bridge: `Physics / Functional Utility / Reliable Simulator RL`. Strict VAM: `No`. Support: `Yes`. Optimization tag: `VAM-RL`. Treats the world model as an imperfect simulator and stabilizes RL through keyframe-initialized rollouts, masked GRPO, and policy-world co-evolution. |
| [ReWorld](https://arxiv.org/abs/2601.12428) | ReWorld: Multi-Dimensional Reward Modeling for Embodied World Models | arXiv 2026 | 1/2026 | `robot-only; exo; action-labeled; scripted` | - | Bridge: `Physics / Functional Utility / RL Alignment`. Strict VAM: `No`. Optimization tag: `VAM-RL`. Uses a multi-dimensional preference reward to post-train embodied world models toward physical realism, task logic, and embodiment plausibility. |
| [Interactive World Simulator](https://arxiv.org/abs/2603.08546) | Interactive World Simulator for Robot Policy Training and Evaluation | arXiv 2026 | 3/2026 | `robot-only; multi-view; action-labeled; scripted` | [Project](https://www.yixuanwang.me/interactive_world_sim/) / [Code](https://github.com/WangYixuan12/interactive_world_sim) | Bridge: `World Model / Interactive Latent Simulator`. Strict VAM: `Partial`. Support: `Yes`. Uses consistency models for image decoding and latent dynamics to build a fast long-horizon simulator for policy training and evaluation. |
| [EVA](https://arxiv.org/abs/2603.17808) | EVA: Aligning Video World Models with Executable Robot Actions via Inverse Dynamics Rewards | arXiv 2026 | 3/2026 | `robot-only; exo; action-labeled; scripted` | [Project](https://eva-project-page.github.io/) / [Code](https://github.com/RobbinW/EVA) | Bridge: `Physics / Functional Utility / Inverse-Dynamics Reward`. Strict VAM: `Partial`. Optimization tag: `VAM-RL`. Uses IDM-derived executability rewards and Flow-GRPO to align a staged video world model toward physically executable futures before action decoding. |
| [World Action Verifier](https://arxiv.org/abs/2604.01985) | World Action Verifier: Self-Improving World Models via Forward-Inverse Asymmetry | arXiv 2026 | 4/2026 | `co-training; exo; action-free + action-labeled; scripted` | [Project](https://world-action-verifier.github.io/) / [Code](https://github.com/world-action-verifier/wav_minigrid) | Bridge: `Physics / Functional Utility / Verification`. Strict VAM: `No`. Decomposes world-model validation into state plausibility and action reachability to improve data acquisition and downstream policy learning. |
| [JailWAM](https://arxiv.org/abs/2604.05498) | JailWAM: Jailbreaking World Action Models in Robot Control | arXiv 2026 | 4/2026 | `robot-only; exo; action-labeled; scripted` | [Project](https://jailwam.github.io/) | Bridge: `Safety / Jailbreak Evaluation`. Strict VAM: `No`. Defines WAM jailbreak attack and evaluation protocols with trajectory-risk mapping and dual-path verification, exposing physical safety failures that visual plausibility metrics miss. |
| [ViVa](https://arxiv.org/abs/2604.08168) | ViVa: A Video-Generative Value Model for Robot Reinforcement Learning | arXiv 2026 | 4/2026 | `robot-only; exo; action-labeled; scripted` | [Project](https://viva-value-model.github.io/) / [Code](https://github.com/GigaAI-research/ViVa) | Bridge: `Physics / Functional Utility / Video-Generative Value`. Strict VAM: `No`. Optimization tag: `VAM-RL`. Rewrites value estimation as future-dynamics prediction with a video generator, giving robot RL a world-model-aware value signal. |
| [RoboWM-Bench](https://arxiv.org/abs/2604.19092) | RoboWM-Bench: A Benchmark for Evaluating World Models in Robotic Manipulation | arXiv 2026 | 4/2026 | `robot-only; exo; action-labeled; scripted` | [Code](https://github.com/fffstrong/RoboWM-Bench) | Bridge: `Evaluation / Embodied Executability`. Strict VAM: `No`. Evaluates whether generated robot videos can be translated back into executable actions and succeed in simulated manipulation, closing the loop between visual futures and control utility. |
| [Hi-WM](https://arxiv.org/abs/2604.21741) | Hi-WM: Human-in-the-World-Model for Scalable Robot Post-Training | arXiv 2026 | 4/2026 | `robot-only; exo; action-labeled; scripted` | [Project](https://hi-wm.github.io/) | Bridge: `World Model / Human Corrective Post-Training`. Strict VAM: `No`. Support: `Yes`. Uses an action-conditioned world model as a rollback-and-branch substrate for human correction, scaling post-training data without requiring every fix to start from a physical reset. |
| [dWorldEval](https://arxiv.org/abs/2604.22152) | dWorldEval: Scalable Robotic Policy Evaluation via Discrete Diffusion World Model | arXiv 2026 | 4/2026 | `robot-only; exo; action-labeled; scripted` | [Project](https://dworldeval.github.io/) | Bridge: `Evaluation / Discrete Diffusion World Model`. Strict VAM: `No`. Tokenizes observations, language, actions, and progress for scalable imagined policy evaluation, extending the evaluation line beyond visual fidelity toward task progress and action controllability. |
| [EA-WM](https://arxiv.org/abs/2605.06192) | EA-WM: Event-Aware Generative World Model with Structured Kinematic-to-Visual Action Fields | arXiv 2026 | 5/2026 | `robot-only; exo; action-labeled; scripted` | - | Bridge: `World Model / Kinematic-to-Visual Action Field`. Strict VAM: `No`. Renders low-dimensional actions into camera-view motion fields and uses event-aware fusion to improve action-conditioned robot video rollouts. |
| [When to Trust Imagination](https://arxiv.org/abs/2605.06222) | When to Trust Imagination: Adaptive Action Execution for World Action Models | arXiv 2026 | 5/2026 | `robot-only; exo; action-labeled; scripted` | - | Bridge: `Execution / Future-Reality Verification`. Strict VAM: `No`. Support: `Yes`. Uses future-reality consistency to decide whether a WAM action chunk remains trustworthy or should be replanned. |
| [Is the Future Compatible?](https://arxiv.org/abs/2605.07514) | Is the Future Compatible? Diagnosing Dynamic Consistency in World Action Models | arXiv 2026 Technical Report | 5/2026 | `robot-only; exo; action-labeled; scripted` | - | Bridge: `Evaluation / Action-State Consistency`. Strict VAM: `No`. Defines action-state compatibility and consensus selection to test whether imagined futures are dynamically compatible with candidate actions. |
| [RoboAlign-R1](https://arxiv.org/abs/2605.03821) | RoboAlign-R1: Distilled Multimodal Reward Alignment for Robot Video World Models | arXiv 2026 | 5/2026 | `robot-only; exo; action-labeled; scripted` | [Code](https://github.com/Alexander-wu/RoboAlign_R1) | Bridge: `Physics / Functional Utility / Multimodal Reward Alignment`. Strict VAM: `No`. Optimization tag: `VAM-RL`. Distills a multimodal robot-video judge into a GRPO reward and uses sliding-window re-encoding to reduce long-rollout drift. |
| [DreamAvoid](https://arxiv.org/abs/2605.11750) | DreamAvoid: Critical-Phase Test-Time Dreaming to Avoid Failures in VLA Policies | arXiv 2026 | 5/2026 | `robot-only; multi-view; action-labeled; scripted` | [Code](https://github.com/XianzheFan/DreamAvoid) | Bridge: `Execution / Test-Time Dreaming`. Strict VAM: `No`. Support: `Yes`. Triggers short-horizon world-model dreaming at critical phases, scores candidate actions with a failure-aware evaluator, and avoids irreversible local mistakes. |
| [Learning Action Manifold](https://arxiv.org/abs/2605.11832) | Learning Action Manifold with Multi-view Latent Priors for Robotic Manipulation | arXiv 2026 | 5/2026 | `robot-only; multi-view; action-labeled; scripted` | [Project](https://junjxiao.github.io/Multi-view-VLA.github.io/) / [Code](https://github.com/junjxiao/Multi-view-VLA) | Bridge: `Explicit 3D / Multi-view Latent Prior + Action Manifold`. Strict VAM: `No`. Support: `Yes`. Uses generated multi-view latent priors and geometry-guided gated fusion to stabilize spatial perception, while action-manifold learning makes the generative action head predict clean action samples instead of unstructured velocity/noise targets. |
| [Feedback World Model](https://arxiv.org/abs/2605.15705) | Feedback World Model Enables Precise Guidance of Diffusion Policy | arXiv 2026 | 5/2026 | `robot-only; exo; action-labeled; scripted` | [Project](https://lorenzo-0-0.github.io/Feedback_World_Model/) | Bridge: `Execution / Feedback-Corrected Latent World Model`. Strict VAM: `No`. Support: `Yes`. Converts a frozen latent world model from open-loop guidance into an observer-like closed-loop test-time guidance module by correcting predicted latent dynamics with real observations and action-aware controllability weights. |
| [RoboFlow4D](https://arxiv.org/abs/2605.17522) | RoboFlow4D: A Lightweight Flow World Model Toward Real-Time Flow-Guided Robotic Manipulation | arXiv 2026 | 5/2026 | `robot-only; exo; action-labeled; scripted` | [Project](https://simonlinsx.github.io/RoboFlow4D_Page/) | Bridge: `Explicit 3D / Flow World Model`. Strict VAM: `Partial`. Support: `Yes`. Predicts lightweight multi-frame 3D flow plans as a low-frequency world-model signal, then conditions fast DP/DiT executors for real-time flow-guided manipulation. |
| [GaussianDream](https://arxiv.org/abs/2605.20752) | GaussianDream: A Feed-Forward 3D Gaussian World Model for Robotic Manipulation | arXiv 2026 | 5/2026 | `robot-only; multi-view; action-labeled; scripted` | [Code](https://github.com/TuojingAI/GaussianDream) | Bridge: `Explicit 3D / 3D Gaussian World Model Prefix`. Strict VAM: `No`. Support: `Yes`. Uses current and future 3D Gaussian reconstruction/prediction as dense training-time supervision, then discards the heavy heads and deploys only a compact geometry-and-dynamics prefix for action generation. |
| [Beyond Euclidean Proximity](https://arxiv.org/abs/2605.22164) | Beyond Euclidean Proximity: Repairing Latent World Models with Horizon-Matched Trajectory Reachability Metrics | arXiv 2026 | 5/2026 | `robot-only; exo; action-labeled; scripted` | - | Bridge: `Evaluation / Latent Reachability Metric`. Strict VAM: `No`. Replaces raw latent Euclidean terminal cost with a horizon-matched trajectory reachability metric, exposing and repairing the planner-facing mismatch between latent proximity and real action reachability. |
| [GEM-4D](https://arxiv.org/abs/2605.22882) | GEM-4D: Geometry-Enhanced Video World Models for Robot Manipulation | arXiv 2026 | 5/2026 | `robot-only; exo; action-labeled; scripted` | [Project](https://anonymous-submission-20.github.io/gem.github.io/) | Bridge: `Explicit 3D / Geometry-Enhanced Video Rollout`. Strict VAM: `Partial`. Distills dense 4D correspondence supervision into a video world model so generated rollouts preserve geometry well enough for inverse-dynamics action extraction. |
| [HumanEgo](https://arxiv.org/abs/2605.24934) | HumanEgo: Zero-Shot Robot Learning from Minutes of Human Egocentric Videos | arXiv 2026 | 5/2026 | `human-ego; action-free; unscripted` | [Project](https://humanego-ai.github.io/) / [Code](https://github.com/TX-Leo/HumanEgo) | Bridge: `Video-to-Robot / Interaction-Centric Human Video`. Strict VAM: `No`. Boundary case: turns minutes of human egocentric hand-object interaction video into robot-executable policies through interaction-centric tokens and dense future-state objectives. |
| [GE-Sim 2.0](https://arxiv.org/abs/2605.27491) | GE-Sim 2.0: A Roadmap Towards Comprehensive Closed-loop Video World Simulators for Robotic Manipulation | arXiv 2026 | 5/2026 | `robot-only; multi-view; action-labeled; scripted` | [Project](https://ge-sim-v2.github.io/) / [Code](https://github.com/AgibotTech/GE-Sim-V2) | Bridge: `World Model / Closed-Loop Video Simulator`. Strict VAM: `Partial`. Support: `Yes`. Adds proprioceptive state prediction, world-judge rewards, and accelerated rollouts so action-conditioned video world models can serve as closed-loop robot simulators. |
| [DexWM](https://arxiv.org/abs/2512.13644) | World Models for Learning Dexterous Hand-Object Interactions from Human Videos | arXiv 2025 | 12/2025 | `co-training; ego; explicit-3D-labeled + action-labeled; unscripted` | [Project](https://raktimgg.github.io/dexwm/) / [Code](https://github.com/facebookresearch/dexwm) | Bridge: `Explicit 3D / Dexterous Hand Keypoints`. Strict VAM: `Partial`. Learns dexterous interaction dynamics from human videos using 3D hand keypoints and hand-consistency losses, then transfers to robot manipulation. |
| [A0](https://arxiv.org/abs/2504.12636) | A0: An Affordance-Aware Hierarchical Model for General Robotic Manipulation | arXiv 2025 | 4/2025 | `robot-only; exo; action-labeled; scripted` | [Project](https://a-embodied.github.io/A0/) / [Code](https://github.com/A-embodied/A0) / [Ckpt](https://huggingface.co/collections/JianZhangAI/a0) | Bridge: `Explicit 2D / Object-Centric Affordance`. Strict VAM: `No`. Object-centric affordance interface for general manipulation. |
| [SKIP](https://arxiv.org/abs/2606.00664) | SKIP: Sparse Keyframe Interpolation Paradigm for Efficient Embodied World Models | arXiv 2026 | 6/2026 | `robot-only; exo; action-labeled; scripted` | - | Bridge: `World Model / Sparse Keyframe Data Synthesis`. Strict VAM: `No`. Support: `Yes`. Uses sparse keyframe embodied world modeling to synthesize policy-training data, so it is better read as R4 support than as a strict VAM method. |
| [Beyond Task Success](https://arxiv.org/abs/2606.01095) | Beyond Task Success: Behavioral and Representational Diagnostics for WAM and VLA | arXiv 2026 | 6/2026 | `evaluation; WAM + VLA diagnostics` | - | Bridge: `Evaluation / Behavior-Representation Diagnostics`. Strict VAM: `No`. Evaluation paper focused on diagnosing behavioral and representation differences between WAMs and VLAs beyond binary success rates. |
| [RoboDream](https://arxiv.org/abs/2606.02577) | RoboDream: Compositional World Models for Scalable Robot Data Synthesis | arXiv 2026 | 6/2026 | `robot-only; exo; action-labeled; scripted` | [Project](https://junjieye.com/RoboDream/) / [Code](https://github.com/Jay-Ye/RoboDream) | Bridge: `World Model / Compositional Robot Data Synthesis`. Strict VAM: `No`. Support: `Yes`. Uses compositional world models to generate scalable robot data, a useful data-synthesis support route rather than a deploy-time VAM. |
| [PiL-World](https://arxiv.org/abs/2606.05773) | PiL-World: A Chunk-Wise World Model for VLA Policy-in-the-Loop Evaluation | arXiv 2026 | 6/2026 | `robot-only; exo; action-labeled; scripted` | - | Bridge: `Evaluation / Policy-in-the-Loop World Model`. Strict VAM: `No`. Builds chunk-wise world-model evaluation around an actual VLA policy loop, directly testing whether imagined futures preserve policy-relevant behavior. |
| [PRISM](https://arxiv.org/abs/2606.07974) | PRISM: PRior-guided Imagination Sampling in world Models | arXiv 2026 | 6/2026 | `robot-only; exo; action-labeled; scripted` | [Project](https://yuhaiw.github.io/PRISM_web/) | Bridge: `Planning / Prior-Guided Imagination Sampling`. Strict VAM: `No`. Support: `Yes`. Samples world-model imagination under prior guidance for planning support rather than proposing a new VAM backbone. |
| [Video2Sim2Real](https://arxiv.org/abs/2606.08828) | Video2Sim2Real: Full-Stack Autonomous Dexterous Skill Acquisition from a Single Human Video | arXiv 2026 | 6/2026 | `human-video; dexterous; action-free-to-robot; scripted` | [Project](https://video2sim2real.github.io/) | Bridge: `Video-to-Robot / Single Human Video Digital-Twin Bridge`. Strict VAM: `No`. Boundary case: full-stack single-human-video to dexterous robot skill pipeline, relevant as a video-to-robot bridge but not a canonical WAM. |
| [Targeting World Models](https://arxiv.org/abs/2606.09499) | Targeting World Models to Compromise Robot Learning Pipelines | arXiv 2026 | 6/2026 | `safety; robot world-model pipeline` | - | Bridge: `Safety / World-Model Pipeline Attack`. Strict VAM: `No`. Safety evaluation paper for attacks against world-model-based robot learning pipelines. |
| [iMaC](https://arxiv.org/abs/2606.09813) | iMaC: Translating Actions into Motion and Contact Images for Embodied World Models | arXiv 2026 | 6/2026 | `robot-only; exo; action-labeled; scripted` | [Code](https://github.com/imac-wm/iMac) | Bridge: `Explicit 2D / Motion and Contact Images`. Strict VAM: `Partial`. Converts actions into structured motion/contact image signals, making it a clean explicit-2D bridge for embodied world models. |
| [EA-WM](https://arxiv.org/abs/2606.13053) | EA-WM: Event-Aware World Models with Task-Specification Grounding for Long-Horizon Manipulation | arXiv 2026 | 6/2026 | `robot-only; exo; action-labeled; scripted` | - | Bridge: `Evaluation / Event-Aware Task-Specification Grounding`. Strict VAM: `No`. New EA-WM distinct from the older `2605.06192` entry; focuses on task-specification-grounded event awareness for long-horizon evaluation/planning. |
| [WEAVER](https://arxiv.org/abs/2606.13672) | WEAVER, Better, Faster, Longer: An Effective World Model for Robotic Manipulation | arXiv 2026 | 6/2026 | `robot-only; exo; action-labeled; scripted` | [Project](https://arnavkj1995.github.io/WEAVER/) / [Code](https://github.com/arnavkj1995/WEAVER) / [Ckpt](https://huggingface.co/arnavkj1995/WEAVER) | Bridge: `World Model / Learned Simulator for Planning`. Strict VAM: `Partial`. Support: `Yes`. Callable learned world model for policy evaluation, improvement, and test-time planning, so it belongs in R4 even though it can support action selection. |
| [WAM-RL](https://arxiv.org/abs/2606.17906) | WAM-RL: World-Action Model Reinforcement Learning with Reconstruction Rewards and Online Video SFT | CVPR Workshops 2026 | 6/2026 | `robot-only; exo; action-labeled + online RL; scripted` | - | Bridge: `Optimization / Online WAM Reinforcement Learning`. Strict VAM: `No`. Support: `Yes`. Optimization tag: `VAM-RL`. Jointly improves actor and world model with reconstruction rewards and online video SFT, targeting recovery and long-horizon correction beyond imitation. |
| [SC3-Eval](https://arxiv.org/abs/2606.18610) | SC3-Eval: Evaluating Robot Foundation Models via Self-Consistent Video Generation | arXiv 2026 | 6/2026 | `evaluation; multi-view; action-labeled; real-policy rollouts` | [Project](https://weichengtseng.github.io/sc3-eval/) | Bridge: `Evaluation / Self-Consistent Video Policy Rollout`. Strict VAM: `No`. Enforces forward-inverse dynamics, cross-view consistency, and test-time consistency so video rollouts better match real policy failures and success rates. |
| [SafeDojo](https://arxiv.org/abs/2606.20698) | SafeDojo: Safe Reinforcement Learning for VLA via Interactive World Model | arXiv 2026 | 6/2026 | `robot-only; exo; action-labeled + safe RL; scripted` | - | Bridge: `Safety / Interactive World-Model Safe RL`. Strict VAM: `No`. Safety: `Yes`. Optimization tag: `VAM-RL`. Uses imagined rollouts, task reward heads, safety cost heads, and Lagrangian GRPO to train safer VLA behavior without dangerous real exploration. |
| [World Value Models](https://arxiv.org/abs/2606.24742) | World Value Models for Robotic Manipulation | arXiv 2026 | 6/2026 | `robot-only; exo; mixed-quality trajectories; scripted` | [Project](https://zh1hao.wang/wvm/) / [Code](https://github.com/255isWhite/WVM) | Bridge: `Value / World-Model Value Curves`. Strict VAM: `No`. Uses video-world-model dynamics to estimate non-monotonic value chunks over hesitation, retry, and recovery trajectories, making value estimation more temporally grounded. |
| [REGEN](https://arxiv.org/abs/2606.27374) | World Action Models Enable Continual Imitation Learning with Recurrent Generative Replays | arXiv 2026 | 6/2026 | `continual; robot-only; generated replay; action-labeled` | [Project](https://manishgovind.github.io/REGEN/) / [Code](https://github.com/ManishGovind/REGEN) | Bridge: `Continual / Recurrent Generative Replay`. Strict VAM: `No`. Support: `Yes`. Uses WAM-generated pseudo-replay trajectories so policies can rehearse old tasks during continual imitation learning without storing original demonstrations. |
| [PhysisForcing](https://arxiv.org/abs/2606.28128) | PhysisForcing: Physics Reinforced World Simulator for Robotic Manipulation | arXiv 2026 | 6/2026 | `robot-only; exo; action-conditioned video; scripted` | [Project](https://dagroup-pku.github.io/PhysisForcing.github.io/) / [Code](https://github.com/DAGroup-PKU/PhysisForcing) | Bridge: `Physics / Hidden-Space Trajectory and Relation Alignment`. Strict VAM: `No`. Support: `Yes`. Adds point-trajectory and semantic-relation alignment on physics-informative regions, improving embodied video simulation and downstream world-model utility. |
| [ICWM](https://arxiv.org/abs/2606.26025) | In-Context World Modeling for Robotic Control | arXiv 2026 | 6/2026 | `robot-only; multi-view; action-labeled + self-probing; scripted` | - | Bridge: `World Model / Interaction-Context System Identification`. Strict VAM: `No`. Support: `Yes`. Uses task-agnostic `(start image, action, end image)` probing prefixes to infer current camera/embodiment configuration before action prediction, improving OOD viewpoint and morphology robustness without parameter updates. |
| [Structured 4D](https://arxiv.org/abs/2607.01166) | Structured 4D Latent Predictive Model for Robot Planning | ICML 2026 | 7/2026 | `robot-only; multi-view RGB-D; explicit-3D-latent; action-labeled; scripted` | [Project](https://structured-4d-model.github.io/) / [Code](https://github.com/sky-lzy/Structured-4D-Model) / [Ckpt](https://huggingface.co/zhiyi24/structured-4d-model) | Bridge: `Explicit 3D / Structured 4D Latent Subgoal`. Strict VAM: `Partial`. Predicts structured future 3D latents that can be decoded into point clouds or 3D Gaussians, then uses inverse dynamics or geometric registration for robot planning. |
| [RoboWorld](https://arxiv.org/abs/2607.01060) | RoboWorld: Fast and Reliable Neural Simulators for Generalist Robot Policy Evaluation | arXiv 2026 | 7/2026 | `evaluation; multi-view; action-labeled; real-policy rollouts` | - | Bridge: `Evaluation / Neural Simulator`. Strict VAM: `No`. Uses Step Forcing and task-progress-aware VLM scoring to turn action-conditioned video rollouts into a fast proxy evaluator whose policy ranking tracks real-world robot results. |
| [PhysMani](https://arxiv.org/abs/2607.01938) | PhysMani: Physics-principled 3D World Model for Dynamic Object Manipulation | ECCV 2026 | 7/2026 | `robot-only; multi-view RGB-D; explicit-3D-Gaussian; dynamic-object; action-labeled; scripted` | [Code](https://github.com/vLAR-group/PhysMani) | Bridge: `Explicit 3D / Physics-Guided 3D Gaussian Dynamics`. Strict VAM: `Partial`. Learns physics-principled 3D Gaussian dynamics and exposes velocity/geometry signals that policies can use for moving-object manipulation. |
| [DynaWM](https://arxiv.org/abs/2607.02604) | DynaWM: A Base-VLA-Guided World Foundation Model for Moving-Object Manipulation | arXiv 2026 | 7/2026 | `robot-only; multi-view; action-labeled; simulated dynamic manipulation` | - | Bridge: `Physical Intelligence / Dynamic Action Regeneration`. Strict VAM: `No`. Support: `Yes`. Treats a frozen VLA action chunk as a cross-model physical interface, then uses multi-view motion history and proprioception to regenerate interception-aware actions. |
| [GigaWorld-1](https://arxiv.org/abs/2607.02642) | GigaWorld-1: A Roadmap to Build World Models for Robot Policy Evaluation | arXiv 2026 | 7/2026 | `mixed robot + human-ego; multi-view; action-labeled; policy rollouts` | [Project](https://open-gigaai.github.io/giga-world-1/) / [Code](https://github.com/open-gigaai/giga-world-1) | Bridge: `Evaluation / Policy-Evaluator Roadmap`. Strict VAM: `No`. Builds WMBench/WMES and a partially open toolchain around long-horizon action-faithful world-model rollouts, emphasizing closed-loop outcome alignment over short-term video fidelity. |
| [WorldSample](https://arxiv.org/abs/2607.02431) | WorldSample: Closed-loop Real-robot RL with World Modelling | arXiv 2026 | 7/2026 | `robot-only; real-robot; action-labeled + online RL; mixed real/synthetic` | [Project](https://xxreinsno.github.io/worldsample/) | Bridge: `Optimization / Real-Synthetic World-Model RL`. Strict VAM: `No`. Support: `Yes`. Optimization tag: `VAM-RL`. Uses real rollouts as grounding, world-model counterfactual trajectories as synthetic replay, and Policy-Paced Learning to reduce model-bias damage during real-robot RL. |
| [RynnWorld-Teleop](https://arxiv.org/abs/2607.06558) | RynnWorld-Teleop: An Action-Conditioned World Model for Digital Teleoperation | arXiv 2026 | 7/2026 | `co-training; ego; depth-aware-hand-pose; action-labeled; teleop-generated` | [Project](https://alibaba-damo-academy.github.io/RynnWorld-Teleop.github.io/) / [Code](https://github.com/alibaba-damo-academy/RynnWorld-Teleop) / [Ckpt](https://huggingface.co/Alibaba-DAMO-Academy/RynnWorld-Teleop) | Bridge: `Video-to-Robot / Digital Teleoperation Action Bridge`. Strict VAM: `No`. Support: `Yes`. Uses an action-conditioned world model plus depth-aware hand-pose and retargeted action signals to synthesize digital-teleoperation data for robot policy learning. |
| [RynnWorld-4D](https://arxiv.org/abs/2607.06559) | RynnWorld-4D: 4D Embodied World Models for Robotic Manipulation | arXiv 2026 | 7/2026 | `co-training; ego; pseudo-RGB-DF; action-free + action-labeled; scripted` | [Project](https://alibaba-damo-academy.github.io/RynnWorld-4D.github.io/) / [Code](https://github.com/alibaba-damo-academy/RynnWorld-4D) / [Ckpt](https://huggingface.co/Alibaba-DAMO-Academy/RynnWorld-4D) | Bridge: `Explicit 3D/4D / Projective RGB-DF Latent`. Strict VAM: `Partial`. Predicts RGB-depth-flow 4D embodied futures and connects the internal 4D latent to action policies for geometry-aware manipulation. |
| [WAM-TTT](https://arxiv.org/abs/2607.06988) | WAM-TTT: Steering World-Action Models by Watching Human Play at Test Time | arXiv 2026 | 7/2026 | `human-video + robot; action-free human video + action-labeled robot; real household` | - | Bridge: `Video-to-Robot / Human-Video Test-Time Memory`. Strict VAM: `Partial`. Support: `Yes`. Writes raw human videos into video-side fast-weight memory inside a frozen WAM via test-time training, letting robot action queries read task-specific memory without robot demonstrations. |

<a id="bridge-signal-view"></a>
# 🌉 Bridge Signal View

🔀 This second view reorganizes the same paper set by executable interface rather than by route.

This view is aligned with the original survey-style bridge taxonomy: `Predictive World Model`, `Latent Action`, `Explicit 2D`, and `Explicit 3D`. In practice, many predictive-world-model papers instantiate that bridge as `video plan`, `future latent`, or `joint video-action` modeling.

## 🌍 Predictive World Model / Video Plan / Joint Video-Action

🌍 Read this bucket when the bridge itself is a predicted future consequence: explicit video plan, future latent, rollout chunk, or joint video-action consequence model.

| Paper | Venue | Route | Strict | Data tags | Project | Code | Checkpoint |
| --- | --- | --- | --- | --- | --- | --- | --- |
| [Dreamitate](https://arxiv.org/abs/2406.16862) | CoRL 2024 | R1 | Partial | `robot-only; exo; action-labeled; scripted` | [site](https://dreamitate.cs.columbia.edu/) | [code](https://github.com/cvlab-columbia/dreamitate) | - |
| [Gen2Act](https://arxiv.org/abs/2409.16283) | arXiv 2024 | R1 | Yes | `co-training; exo; action-free + action-labeled; scripted` | [site](https://homangab.github.io/gen2act/) | - | - |
| [UniPi](https://arxiv.org/abs/2302.00111) | NeurIPS 2023 | R1 | Partial | `robot-only; exo; action-labeled; scripted` | - | [code](https://github.com/universal-policy/unipi) | - |
| [GR-1](https://arxiv.org/abs/2312.13139) | ICLR 2024 | Support / R1-R2 | Partial | `robot-only; exo; action-labeled; scripted` | [site](https://gr1-manipulation.github.io/) | [code](https://github.com/bytedance/GR-1) | - |
| [VILP](https://arxiv.org/abs/2502.01784) | arXiv 2025 | R1 | Yes | `robot-only; multi-view; action-labeled; scripted` | - | [code](https://github.com/ZhengtongXu/VILP) | - |
| [Grounding Video Models to Actions through Goal Conditioned Exploration](https://arxiv.org/abs/2411.07223) | ICLR 2025 Spotlight | R1 | Partial | `robot-only; exo; action-free + action-labeled; scripted` | - | - | - |
| [VPP](https://arxiv.org/abs/2412.14803) | ICML 2025 Spotlight | R1 | Yes | `robot-only; exo; action-labeled; scripted` | [site](https://video-prediction-policy.github.io/) | [code](https://github.com/roboterax/video-prediction-policy) | [ckpt](https://huggingface.co/yjguo/svd-robot-calvin-ft) |
| [Seer](https://arxiv.org/abs/2412.15109) | ICLR 2025 | R1 | Partial | `robot-only; exo; action-labeled; scripted` | [site](https://nimolty.github.io/Seer/) | [code](https://github.com/InternRobotics/Seer) | - |
| [FLARE](https://arxiv.org/abs/2505.15659) | CoRL 2025 | R1 / R3 (boundary) | Partial | `co-training; ego; action-free + action-labeled; scripted` | [site](https://research.nvidia.com/labs/gear/flare/) | - | - |
| [FRAPPE](https://arxiv.org/abs/2602.17259) | arXiv 2026 | Support / R1 | Partial | `co-training; exo; action-free + action-labeled; scripted` | [site](https://h-zhao1997.github.io/frappe) | [code](https://github.com/OpenHelix-Team/frappe) | [ckpt](https://huggingface.co/collections/hhhJB/frappe) |
| [DreamGen](https://arxiv.org/abs/2505.12705) | arXiv 2025 | Support / R1 | No | `robot-only; exo; action-labeled; scripted` | - | - | - |
| [LaDi-WM](https://arxiv.org/abs/2505.11528) | CoRL 2025 | R1 | Partial | `robot-only; exo; action-labeled; scripted` | [site](https://guhuangai.github.io/LaDiWM.github.io/) | [code](https://github.com/GuHuangAI/LaDiWM) | - |
| [DreamDojo](https://arxiv.org/abs/2602.06949) | arXiv 2026 | Support / R1 | Partial | `co-training; ego; action-free + action-labeled; scripted` | [site](https://dreamdojo-world.github.io/) | - | - |
| [Say, Dream, and Act](https://arxiv.org/abs/2602.10717) | arXiv 2026 | R1 | Yes | `robot-only; exo; action-labeled; scripted` | - | - | - |
| [DiT4DiT](https://arxiv.org/abs/2603.10448) | arXiv 2026 | R1 | Yes | `robot-only; ego; action-labeled; scripted` | [site](https://dit4dit.github.io/) | - | - |
| [FutureVLA](https://arxiv.org/abs/2603.10712) | arXiv 2026 | R1 | Partial | `robot-only; exo; action-labeled; scripted` | [site](https://xuxiaoxxxx.github.io/projects/futurevla/) | [code](https://github.com/xuxiaoxxxx/FutureVLA) | - |
| [DIAL](https://arxiv.org/abs/2603.29844) | arXiv 2026 | R1 | Yes | `co-training; exo; action-free + action-labeled; scripted` | - | - | - |
| [S-VAM](https://arxiv.org/abs/2603.16195) | arXiv 2026 | R1 | Yes | `robot-only; exo; action-labeled; scripted` | [site](https://haodong-yan.github.io/S-VAM/) | [code](https://github.com/Haodong-Yan/S-VAM-Code) | - |
| [StructVLA](https://arxiv.org/abs/2603.12553) | arXiv 2026 | R1 | Partial | `robot-only; exo; action-labeled; scripted` | - | - | - |
| [VAMPO](https://arxiv.org/abs/2603.19370) | arXiv 2026 | R1 | Yes | `robot-only; exo; action-labeled; scripted` | [site](https://vampo-robot.github.io/VAMPO/) | [code](https://github.com/OpenHelix-Team/VAMPO) | [ckpt](https://huggingface.co/williammmgezju/VAMPO_SVD) |
| [Veo-Act](https://arxiv.org/abs/2604.04502) | arXiv 2026 | R1 | Partial | `robot-only; multi-view; action-labeled; scripted` | - | - | - |
| [mimic-video](https://arxiv.org/abs/2512.15692) | arXiv 2025 | R1 | Yes | `robot-only; exo; action-labeled; scripted` | [site](https://mimic-video.github.io/) | - | - |
| [Vidar](https://arxiv.org/abs/2507.12898) | arXiv 2025 | R1 | Partial | `robot-only; exo; action-labeled; scripted` | [site](https://embodiedfoundation.github.io/vidar_anypos) | [code](https://github.com/thu-ml/vidar) | [ckpt](https://huggingface.co/Xiang-cd/vidar) |
| [Vidarc](https://arxiv.org/abs/2512.17661) | arXiv 2025 | R1 | Partial | `robot-only; exo; action-labeled; scripted` | [site](https://embodiedfoundation.github.io/vidar_anypos) | [code](https://github.com/thu-ml/vidar) | [ckpt](https://huggingface.co/Xiang-cd/vidar) |
| [TC-IDM](https://arxiv.org/abs/2601.18323) | arXiv 2026 | R1 | Partial | `robot-only; exo; action-labeled; scripted` | [site](https://wsbaiyi.github.io/TC-IDM-Page/) | [code](https://github.com/wsbaiyi/TC-IDM) | - |
| [DVA](https://www.rhoda.ai/research/direct-video-action) | Rhoda AI Research 2026 | R1 | Yes | `co-training; exo; action-free + action-labeled; unscripted` | [site](https://www.rhoda.ai/research/direct-video-action) | - | - |
| [VERA](https://arxiv.org/abs/2605.27817) | arXiv 2026 | R1 | Yes | `co-training; exo; action-free + action-labeled; unscripted` | [site](https://vera.csail.mit.edu) | - | - |
| [Demo-JEPA](https://arxiv.org/abs/2605.20811) | arXiv 2026 | R1 | Partial | `co-training; exo; action-free + action-labeled; scripted` | [site](https://log2r.github.io/Demo-JEPA/) | [code](https://github.com/banban3forever/Demo-JEPA) | - |
| [StableIDM](https://arxiv.org/abs/2604.17887) | arXiv 2026 | R1 | No | `robot-only; exo; action-labeled; scripted` | - | - | - |
| [T3VF](https://arxiv.org/abs/2605.08215) | arXiv 2026 | Support / R1-R4 | No | `robot-only; exo; action-labeled; scripted` | - | - | - |
| [PAD](https://arxiv.org/abs/2411.18179) | NeurIPS 2024 | R2 | Partial | `co-training; exo; action-free + action-labeled; scripted` | - | [code](https://github.com/Robert-gyj/Prediction_with_Action) | [ckpt](https://huggingface.co/yjguo/pad_bridge_pre) |
| [GR-2](https://arxiv.org/abs/2410.06158) | arXiv 2024 | R2 | Partial | `co-training; exo; action-free + action-labeled; scripted` | - | - | - |
| [Video Generators are Robot Policies](https://arxiv.org/abs/2508.00795) | arXiv 2025 | R2 | Partial | `co-training; exo; action-free + action-labeled; scripted` | - | [code](https://github.com/cvlab-columbia/videopolicy) | - |
| [VideoVLA](https://arxiv.org/abs/2512.06963) | NeurIPS 2025 | R2 | Partial | `robot-only; exo; action-labeled; scripted` | [site](https://videovla-nips2025.github.io/) | [code](https://github.com/VideoVLA-Project/VideoVLA) | - |
| [CoVAR](https://arxiv.org/abs/2512.16023) | arXiv 2025 | R2 | Yes | `robot-only; exo; action-labeled; scripted` | - | - | - |
| [Cosmos Policy](https://arxiv.org/abs/2601.16163) | arXiv 2026 | R2 | Partial | `robot-only; exo; action-labeled; scripted` | [site](https://research.nvidia.com/labs/cosmos-lab/cosmos-policy/) | [code](https://github.com/NVlabs/cosmos-policy) | [ckpt](https://huggingface.co/collections/nvidia/cosmos-policy) |
| [DreamZero](https://arxiv.org/abs/2602.15922) | arXiv 2026 | R2 | Partial | `robot-only; exo; action-labeled; scripted` | [site](https://dreamzero0.github.io/) | [code](https://github.com/dreamzero0/DreamZero) | [ckpt](https://huggingface.co/GEAR-Dreams/DreamZero-DROID) |
| [WorldVLA](https://arxiv.org/abs/2506.21539) | arXiv 2025 | R2 | Partial | `robot-only; exo; action-labeled; scripted` | - | [code](https://github.com/alibaba-damo-academy/RynnVLA-002) | [ckpt](https://huggingface.co/Alibaba-DAMO-Academy/WorldVLA) |
| [LingBot-VA](https://arxiv.org/abs/2601.21998) | arXiv 2026 | R2 | Partial | `robot-only; exo; action-labeled; scripted` | [site](https://technology.robbyant.com/lingbot-va) | [code](https://github.com/robbyant/lingbot-va) | [ckpt](https://huggingface.co/collections/robbyant/lingbot-va) |
| [MinD](https://arxiv.org/abs/2506.18897) | arXiv 2025 | R2 / R4 | Partial | `robot-only; exo; action-labeled; scripted` | [site](https://manipulate-in-dream.github.io/) | [code](https://github.com/manipulate-in-dream/MinD) | - |
| [iMoWM](https://arxiv.org/abs/2510.09036) | arXiv 2025 | R2 | Partial | `robot-only; exo; action-labeled; scripted` | [site](https://xingyoujun.github.io/imowm/) | - | - |
| [Ctrl-World](https://arxiv.org/abs/2510.10125) | ICLR 2026 | Support / R4 | Partial | `robot-only; multi-view; action-labeled; scripted` | [site](https://ctrl-world.github.io/) | [code](https://github.com/Robert-gyj/Ctrl-World) | - |
| [VLAW](https://arxiv.org/abs/2602.12063) | arXiv 2026 | R2 | Partial | `robot-only; exo; action-labeled; scripted` | [site](https://sites.google.com/view/vlaw-arxiv) | [code](https://github.com/Robert-gyj/Ctrl-World) | [ckpt](https://huggingface.co/yjguo/Ctrl-World) |
| [World-VLA-Loop](https://arxiv.org/abs/2602.06508) | arXiv 2026 | R2 | Partial | `robot-only; exo; action-labeled; scripted` | [site](https://showlab.github.io/World-VLA-Loop/) | [code](https://github.com/showlab/World-VLA-Loop) | - |
| [Fast-WAM](https://arxiv.org/abs/2603.16666) | arXiv 2026 | R2 | Partial | `robot-only; exo; action-labeled; scripted` | [site](https://yuantianyuan01.github.io/FastWAM/) | [code](https://github.com/yuantianyuan01/FastWAM) | - |
| [Privileged Foresight Distillation](https://arxiv.org/abs/2604.25859) | arXiv 2026 | Support / R2 | Partial | `robot-only; exo; action-labeled; scripted` | - | - | - |
| [GigaWorld-Policy](https://arxiv.org/abs/2603.17240) | arXiv 2026 | R2 / R3 (boundary) | Partial | `robot-only; exo; action-labeled; scripted` | [site](https://gigaai-research.github.io/GigaWorld-Policy/) | [code](https://github.com/open-gigaai/giga-world-policy) | - |
| [World-Value-Action Model](https://arxiv.org/abs/2604.14732) | arXiv 2026 | R2 | Partial | `robot-only; exo; action-labeled; scripted` | - | - | - |
| [STARRY](https://arxiv.org/abs/2604.26848) | arXiv 2026 | R2 | Partial | `robot-only; multi-view; action-labeled; scripted` | - | - | - |
| [VAG](https://arxiv.org/abs/2604.09330) | arXiv 2026 | Support / R2 | Partial | `robot-only; exo; action-labeled; scripted` | - | - | - |
| [OA-WAM](https://arxiv.org/abs/2605.06481) | arXiv 2026 | R2 / R3 (boundary) | Partial | `robot-only; exo; action-labeled; scripted` | - | - | - |
| [HarmoWAM](https://arxiv.org/abs/2605.10942) | arXiv 2026 | R2 / R3 | Partial | `robot-only; exo; action-labeled; scripted` | [site](https://elbb-yu.github.io/HarmoWAM/) | [code](https://github.com/xuanxuanzzzii/HarmoWAM) | - |
| [NoiseGate](https://arxiv.org/abs/2605.07794) | arXiv 2026 | R2 | Partial | `robot-only; exo; action-labeled; scripted` | - | - | - |
| [SANTS](https://arxiv.org/abs/2605.27947) | arXiv 2026 | Support / R2 | No | `robot-only; exo; action-labeled; scripted` | [site](https://advanced-robotics-lab.github.io/SANTS/) | [code](https://github.com/Advanced-Robotics-Lab/SANTS-code) | - |
| [Interactive World Simulator](https://arxiv.org/abs/2603.08546) | arXiv 2026 | Support / R4 | Partial | `robot-only; multi-view; action-labeled; scripted` | [site](https://www.yixuanwang.me/interactive_world_sim/) | [code](https://github.com/WangYixuan12/interactive_world_sim) | - |
| [UWM](https://arxiv.org/abs/2504.02792) | RSS 2025 | R3 | Yes | `robot-only; exo; action-free + action-labeled; scripted` | [site](https://weirdlabuw.github.io/uwm/) | [code](https://github.com/WEIRDLabUW/unified-world-model) | [ckpt](https://drive.google.com/drive/folders/1M4AuVLMRpSwOf_YAp56bV9AqyZI9ul6g?usp=sharing) |
| [UVA](https://arxiv.org/abs/2503.00200) | RSS 2025 | R3 | Partial | `robot-only; exo; action-free + action-labeled; scripted` | [site](https://unified-video-action-model.github.io/) | [code](https://github.com/ShuangLI59/unified_video_action) | [ckpt](https://drive.google.com/file/d/11c2VrmaRp48yw__5A5xpcu8EPzkexHSi/view?usp=sharing) |
| [Being-H0.7](https://research.beingbeyond.com/being-h07) | BeingBeyond 2026 | R3 | Partial | `co-training; ego; action-free + action-labeled; unscripted` | [site](https://research.beingbeyond.com/being-h07) | [code](https://github.com/BeingBeyond/Being-H) | - |
| [MotuBrain](https://arxiv.org/abs/2604.27792) | arXiv 2026 | R3 | Yes | `co-training; multi-view; action-free + action-labeled; scripted` | [site](https://www.shengshu.com/en/motubrain) | - | - |
| [CKT-WAM](https://arxiv.org/abs/2605.06247) | arXiv 2026 | Support / R3 | Partial | `robot-only; exo; action-labeled; scripted` | - | [code](https://github.com/YuhuaJiang2002/CKT-WAM) | - |
| [UWM-JEPA](https://arxiv.org/abs/2605.25313) | arXiv 2026 | Support / R3-R4 | No | `world-model; action-conditioned; latent-belief` | - | [code](https://github.com/santoshkumarradha/uwm-jepa) | - |
| [tau0-WM](https://arxiv.org/abs/2606.01027) | arXiv 2026 | R3 / R2 | Partial | `robot-only; exo; action-labeled; scripted` | [site](https://finch.agibot.com/research/tau0-wm) | [code](https://github.com/sii-research/tau-0-wm) | - |
| [SeeTraceAct](https://arxiv.org/abs/2606.02745) | arXiv 2026 | R1 / R4 | Partial | `co-training; exo; action-free + action-labeled; scripted` | - | [code](https://github.com/jaehyeon-son/SeeTraceAct) | - |
| [WALL-WM](https://arxiv.org/abs/2606.01955) | arXiv 2026 | R2 / R3 | Partial | `robot-only; multi-view; action-labeled; scripted` | - | [code](https://github.com/X-Square-Robot/wall-x) | - |
| [Flash-WAM](https://arxiv.org/abs/2606.05254) | arXiv 2026 | Support / R2 | No | `robot-only; exo; action-labeled; scripted` | [site](https://flashwam.github.io/) | [code](https://github.com/NU-World-Model-Embodied-AI/Flash-WAM) | - |
| [AdaWAM](https://arxiv.org/abs/2606.07089) | arXiv 2026 | R2 / R3 | Partial | `robot-only; exo; action-labeled; scripted` | [site](https://adawam.github.io/) | - | - |
| [AHA-WAM](https://arxiv.org/abs/2606.09811) | arXiv 2026 | R2 / R3 | Partial | `robot-only; exo; action-labeled; scripted` | [site](https://serene-sivy.github.io/aha-wam/) | - | - |
| [Efficient-WAM](https://arxiv.org/abs/2606.10040) | arXiv 2026 | R2 | Partial | `robot-only; exo; action-labeled; scripted` | - | - | - |
| [HiMem-WAM](https://arxiv.org/abs/2606.10363) | arXiv 2026 | R2 / R3 | Partial | `robot-only; exo; action-labeled; scripted` | - | - | - |
| [EWAM](https://arxiv.org/abs/2606.12690) | arXiv 2026 | Support / R2-R4 | No | `robot-only; exo; action-labeled; scripted` | - | - | - |
| [Making Foresight Actionable](https://arxiv.org/abs/2606.12217) | arXiv 2026 | Support / R2 | No | `robot-only; exo; action-labeled; scripted` | - | - | - |
| [World Pilot](https://arxiv.org/abs/2606.12403) | arXiv 2026 | Support / R2 | Partial | `robot-only; exo; action-labeled; scripted` | - | - | - |
| [C3ache](https://arxiv.org/abs/2606.08962) | arXiv 2026 | Support / R2 | No | `robot-only; exo; action-labeled; scripted` | - | - | - |
| [World-Language-Action Model](https://arxiv.org/abs/2606.05979) | arXiv 2026 | R3 | Partial | `robot-only; exo; action-labeled; scripted` | - | [code](https://github.com/SJTU-DENG-Lab/WLA) | - |
| [omega-EVA](https://arxiv.org/abs/2606.09457) | arXiv 2026 | R1 / R4 | Partial | `robot-only; exo; action-labeled; scripted` | - | - | - |
| [MemoryVLA++](https://arxiv.org/abs/2606.09827) | arXiv 2026 | R1 / R4 boundary | Partial | `robot-only; exo; action-labeled; scripted` | - | - | - |
| [ImageWAM](https://arxiv.org/abs/2606.19531) | arXiv 2026 | R1 / R2 | Partial | `robot-only; exo; action-labeled; scripted` | [site](https://zhangwenyao1.github.io/ImageWAM/) | [code](https://github.com/yuyangalin/ImageWAM) | [ckpt](https://huggingface.co/collections/yuyangalin/imagewam) |
| [MemoryVAM](https://arxiv.org/abs/2606.20679) | arXiv 2026 | R1 / R4 | Partial | `robot-only; exo; action-labeled; scripted` | [site](https://memoryvam.github.io/) | - | - |
| [MemoryWAM](https://arxiv.org/abs/2606.20562) | arXiv 2026 | R3 / R4 | Partial | `robot-only; exo; action-labeled; scripted` | [site](https://yangsizhe.github.io/MemoryWAM/) | [code](https://github.com/yangsizhe/MemoryWAM) | - |
| [MV-WAM](https://arxiv.org/abs/2606.21088) | arXiv 2026 | R2 / VAM-RL boundary | Partial | `robot-only; exo; action-labeled; scripted` | - | - | - |
| [DIM-WAM](https://arxiv.org/abs/2606.27677) | arXiv 2026 | R2 / R4 | Partial | `robot-only; exo; action-labeled; scripted` | [site](https://wangkai-casia.github.io/dim-wam/) | [code](https://github.com/wangkai-casia/dim-wam) | - |
| [ICWM](https://arxiv.org/abs/2606.26025) | arXiv 2026 | Support / R4 | No | `robot-only; multi-view; action-labeled + self-probing; scripted` | - | - | - |
| [DVG-WM](https://arxiv.org/abs/2606.32028) | ECCV 2026 | R1 / R4 | Partial | `robot-only; multi-view; action-labeled; scripted` | [site](https://zyshan0929.github.io/DVGWM/) | [code](https://github.com/zyshan0929/DVG-WM) | - |
| [A2World](https://arxiv.org/abs/2606.29501) | ECCV 2026 | R2 / R4 | Partial | `robot-only; multi-view; action-labeled; scripted` | - | [code](https://github.com/LogosRoboticsGroup/A2World) | - |
| [VLAFlow](https://arxiv.org/abs/2607.01586) | arXiv 2026 | Support / R1-R3 | No | `robot-only; exo; action-labeled; scripted` | [site](https://mindvla-team.github.io/VLAFlow/) | [code](https://github.com/MindVLA-Team/VLAFlow) | - |
| [ABot-M0.5](https://arxiv.org/abs/2607.00678) | arXiv 2026 | R3 | Yes | `robot-only; mobile-manipulation; multi-view; action-labeled` | - | [code](https://github.com/amap-cvlab/ABot-Manipulation) | - |
| [DSWAM](https://arxiv.org/abs/2607.04927) | arXiv 2026 | R3 | Partial | `robot-only; multi-view; dual-arm; action-labeled` | - | - | - |
| [InternVLA-A1.5](https://arxiv.org/abs/2607.04988) | arXiv 2026 | R3 / R1 | Partial | `co-training; exo; action-labeled + multimodal; scripted` | [site](https://internrobotics.github.io/internvla-a15.github.io) | [code](https://github.com/InternRobotics/InternVLA-A-series) | [ckpt](https://huggingface.co/collections/InternRobotics/internvla-a15) |
| [VT-WAM](https://arxiv.org/abs/2607.02503) | arXiv 2026 | R2 | Yes | `robot-only; visuotactile; action-labeled; scripted` | - | - | - |
| [WorldBagel](https://arxiv.org/abs/2607.03461) | arXiv 2026 | R3 | Partial | `co-training; exo; action-free + action-labeled; scripted` | - | - | - |
| [Qantara](https://arxiv.org/abs/2607.04978) | arXiv 2026 | R2 / R3 | Partial | `robot-only; exo; action-labeled; scripted` | [site](https://corl-team.github.io/qantara/) | - | - |
| [RoboTALES](https://arxiv.org/abs/2607.06018) | arXiv 2026 | R1 | Partial | `robot-only; exo; action-labeled; scripted` | - | [code](https://github.com/hananshafi/RoboTALES) | - |
| [EgoWAM](https://arxiv.org/abs/2607.08436) | arXiv 2026 | R3 | Partial | `co-training; ego; action-free + action-labeled; unscripted` | [site](https://gatech-rl2.github.io/egowam.github.io/) | - | - |
| [LingBot-VA 2.0](https://arxiv.org/abs/2607.08639) | arXiv 2026 | R2 / R3 | Yes | `co-training; exo; action-free + action-labeled; scripted` | - | - | - |
| [Lumo-2](https://arxiv.org/abs/2607.11270) | arXiv 2026 | R3 | Yes | `robot-only; exo; action-labeled; scripted` | - | - | - |
| [Jetson-PI](https://arxiv.org/abs/2607.12659) | arXiv 2026 | R1 / R3 | Partial | `robot-only; exo; action-labeled; scripted` | - | [code](https://github.com/PKU-SEC-Lab/Jetson-PI) | - |
| [GigaWorld-Policy-0.5](https://arxiv.org/abs/2607.13960) | arXiv 2026 | R3 | Partial | `robot-only; exo; action-labeled; scripted` | [site](https://open-gigaai.github.io/giga-world-policy/) | [code](https://github.com/open-gigaai/giga-world-policy) | - |
| [RxBrain](https://arxiv.org/abs/2607.14187) | arXiv 2026 | R1 / R3 | Partial | `robot-only; exo; action-labeled; scripted` | - | - | - |
| [WorldScape Policy 2.0](https://arxiv.org/abs/2607.18840) | arXiv 2026 | R3 | Partial | `robot-only; exo; action-labeled + multimodal; scripted` | - | - | - |
| [HOST](https://arxiv.org/abs/2607.20033) | arXiv 2026 | R1 | Yes | `co-training; exo; action-free + action-labeled; scripted` | - | - | - |

## 🧩 Latent Action

🧩 Read this bucket when the paper first tries to extract a pseudo-action, motion token, or embodiment-transferable control latent from video, then align it to executable action.

| Paper | Venue | Route | Strict | Data tags | Project | Code | Checkpoint |
| --- | --- | --- | --- | --- | --- | --- | --- |
| [LAPA](https://arxiv.org/abs/2410.11758) | ICLR 2025 | R1 | No | `co-training; exo; action-free + action-labeled; scripted` | [site](https://latentactionpretraining.github.io/) | [code](https://github.com/LatentActionPretraining/LAPA) | [ckpt](https://huggingface.co/latent-action-pretraining/LAPA-7B-openx) |
| [UniVLA](https://arxiv.org/abs/2505.06111) | RSS 2025 | R3 (boundary) | No | `co-training; exo; action-free + action-labeled; scripted` | - | [code](https://github.com/OpenDriveLab/UniVLA) | - |
| [ConLA](https://arxiv.org/abs/2602.00557) | arXiv 2026 | R1 | No | `co-training; exo; action-free + action-labeled; scripted` | - | [code](https://github.com/WeishengDAI/ConLA) | - |
| [DreamDojo](https://arxiv.org/abs/2602.06949) | arXiv 2026 | Support / R1 | Partial | `co-training; ego; action-free + action-labeled; scripted` | [site](https://dreamdojo-world.github.io/) | - | - |
| [Motus](https://arxiv.org/abs/2512.13030) | arXiv 2025 | R3 | Yes | `co-training; exo; action-free + action-labeled; scripted` | [site](https://motus-robotics.github.io/motus) | [code](https://github.com/thu-ml/Motus) | [ckpt](https://huggingface.co/motus-robotics/Motus) |
| [LAWM](https://arxiv.org/abs/2601.05230) | arXiv 2026 | R3 (boundary) | Partial | `co-training; exo; action-free + action-labeled; unscripted` | - | - | - |
| [JALA](https://arxiv.org/abs/2602.21736) | CVPR 2026 | R3 (boundary) | No | `co-training; ego; action-free + action-labeled; unscripted` | - | [code](https://github.com/BeingBeyond/JALA) | - |
| [World2Act](https://arxiv.org/abs/2603.10422) | arXiv 2026 | Support / R1 | No | `robot-only; exo; action-labeled; scripted` | [site](https://wm2act.github.io/) | - | - |
| [MoLA](https://arxiv.org/abs/2605.12167) | ICML 2026 | R1 | Partial | `co-training; exo; action-free + action-labeled; scripted` | [site](https://logosroboticsgroup.github.io/MoLA/) | [code](https://github.com/LogosRoboticsGroup/MoLA) | - |
| [ALAM](https://arxiv.org/abs/2605.10819) | arXiv 2026 | R3 (boundary) | No | `co-training; exo; action-free + action-labeled; scripted` | - | - | - |
| [DiLA](https://arxiv.org/abs/2605.15725) | ICML 2026 | R3 | Partial | `co-training; exo; action-free + action-labeled; scripted` | [site](http://disentangled-latent-action-world-models.github.io) | - | - |
| [HumanEgo](https://arxiv.org/abs/2605.24934) | arXiv 2026 | Support / R4 | No | `human-ego; action-free; unscripted` | [site](https://humanego-ai.github.io/) | [code](https://github.com/TX-Leo/HumanEgo) | - |
| [From Pixels to Tokens](https://arxiv.org/abs/2605.04678) | arXiv 2026 | Survey / Support | No | `systematic-study; latent-action-supervision` | - | [code](https://github.com/RUCKBReasoning/From_Pixels_to_Tokens) | - |
| [EC-Flow](https://arxiv.org/abs/2507.06224) | ICCV 2025 | R3 (boundary) | No | `co-training; exo; action-free + action-labeled; scripted` | [site](https://ec-flow1.github.io/) | [code](https://github.com/YixiangChen515/EC-Flow) | [ckpt](https://huggingface.co/YixiangChen/EC-Flow) |
| [Learning Skills from Action-Free Videos](https://arxiv.org/abs/2512.20052) | arXiv 2025 | R3 (boundary) | No | `co-training; exo; action-free + action-labeled; scripted` | - | - | - |
| [CLAW](https://arxiv.org/abs/2606.04130) | arXiv 2026 | R3 | Partial | `co-training; exo; action-free + action-labeled; scripted` | - | - | - |
| [LARA](https://arxiv.org/abs/2606.07100) | arXiv 2026 | Support / R3 | No | `co-training; exo; action-free + action-labeled; scripted` | - | - | - |
| [RepWAM](https://arxiv.org/abs/2606.13674) | arXiv 2026 | R2 / Latent Action | Partial | `robot-only; exo; action-labeled; scripted` | [site](https://wdrink.github.io/RepWAM/) | [code](https://github.com/wdrink/RepWAM) | - |
| [LaWAM](https://arxiv.org/abs/2606.15768) | arXiv 2026 | R1 / Latent Action | Yes | `co-training; ego; action-free + action-labeled; scripted` | [site](https://rlinf.github.io/LaWAM/) | [code](https://github.com/RLinf/LaWAM) | [ckpt](https://huggingface.co/collections/jialei02/lawam-checkpoints) |
| [WALA](https://arxiv.org/abs/2607.11397) | arXiv 2026 | R3 / Latent Action | Partial | `co-training; exo; action-free + action-labeled; scripted` | [site](https://liujiahao2077.github.io/WALA.github.io/) | [code](https://github.com/liujiahao2077/WALA) | - |

## 📍 Explicit 2D

📍 Read this bucket when the bridge is intentionally sparse and structured in image space: marks, keypoints, tracks, 2D affordances, or embodiment masks.

| Paper | Venue | Route | Strict | Data tags | Project | Code | Checkpoint |
| --- | --- | --- | --- | --- | --- | --- | --- |
| [BridgeV2W](https://arxiv.org/abs/2602.03793) | arXiv 2026 | R2 / R4 | Partial | `robot-only; exo; action-labeled; scripted` | [site](https://bridgev2w.github.io/) | - | - |
| [Multi-View Video Diffusion Policy (MV-VDP)](https://arxiv.org/abs/2604.03181) | arXiv 2026 | R2 | Yes | `robot-only; multi-view; action-labeled; scripted` | [site](https://lpy1219.github.io/MV-VDP-Web/) | - | - |
| [Action Images](https://arxiv.org/abs/2604.06168) | arXiv 2026 | R2 / R3 (boundary) | Yes | `robot-only; multi-view; action-labeled; scripted` | [site](https://actionimages.github.io/) | [code](https://github.com/UMass-Embodied-AGI/ActionImages) | - |
| [AIM](https://arxiv.org/abs/2604.11135) | arXiv 2026 | R2 | Yes | `robot-only; multi-view; action-labeled; scripted` | - | - | - |
| [Mask World Model](https://arxiv.org/abs/2604.19683) | arXiv 2026 | R1 | Partial | `robot-only; multi-view; action-labeled; scripted` | - | [code](https://github.com/LYFCLOUDFAN/mask-world-model) | - |
| [Point Tracking Improves WAM](https://arxiv.org/abs/2605.23856) | arXiv 2026 | R2 / R4 | Partial | `robot-only; exo; action-labeled; scripted` | - | - | - |
| [A0](https://arxiv.org/abs/2504.12636) | arXiv 2025 | R1 / R4 | No | `robot-only; exo; action-labeled; scripted` | [site](https://a-embodied.github.io/A0/) | [code](https://github.com/A-embodied/A0) | [ckpt](https://huggingface.co/collections/JianZhangAI/a0) |
| [MaskWAM](https://arxiv.org/abs/2606.13515) | arXiv 2026 | R2 / Explicit 2D | Partial | `robot-only; exo; action-labeled; scripted` | [site](https://hanyangyu1021.github.io/maskwam.github.io/) | [code](https://github.com/hanyangyu1021/maskwam) | - |
| [iMaC](https://arxiv.org/abs/2606.09813) | arXiv 2026 | R4 / Explicit 2D | Partial | `robot-only; exo; action-labeled; scripted` | - | [code](https://github.com/imac-wm/iMac) | - |
| [Bridge-WA](https://arxiv.org/abs/2607.02195) | arXiv 2026 | R2 / Explicit 2D | Partial | `robot-only; exo; action-labeled; structured future-change maps` | [site](https://hcplab-sysu.github.io/BRIDGE-WA/) | [code](https://github.com/HCPLab-SYSU/BRIDGE-WA) | [ckpt](https://huggingface.co/baiyu858/Bridge-WA-VLABench) |
| [KAM-WM](https://arxiv.org/abs/2607.04652) | arXiv 2026 | R1 / Explicit 2D | Partial | `robot-only; exo; action-labeled; scripted` | - | - | - |
| [FlowWAM](https://arxiv.org/abs/2607.13017) | arXiv 2026 | R2 / Explicit 2D | Yes | `co-training; exo; action-free + action-labeled; scripted` | [site](https://flow-wam.github.io/) | [code](https://github.com/YixiangChen515/FlowWAM) | - |
| [FoMoVLA](https://arxiv.org/abs/2607.14739) | arXiv 2026 | R1 / Explicit 2D | Partial | `robot-only; exo; action-labeled; scripted` | [site](https://liauto-research.github.io/FoMoVLA/) | - | - |
| [Masked Visual Actions](https://arxiv.org/abs/2607.19343) | arXiv 2026 | R1 / R4 / Explicit 2D | Partial | `robot-only; exo; action-labeled; scripted` | [site](https://masked-visual-actions.github.io/) | [code](https://github.com/HadiZayer/masked-visual-actions) | - |

## 🧱 Explicit 3D

🧱 Read this bucket when the bridge is grounded in geometry: hand pose, wrist-fingertip action, object pose, 4D consistency, or other explicit 3D structure.

| Paper | Venue | Route | Strict | Data tags | Project | Code | Checkpoint |
| --- | --- | --- | --- | --- | --- | --- | --- |
| [MVISTA-4D](https://arxiv.org/abs/2602.09878) | arXiv 2026 | R2 / R4 | Partial | `robot-only; multi-view; action-labeled; scripted` | - | - | - |
| [X-WAM](https://arxiv.org/abs/2604.26694) | arXiv 2026 | R3 | Yes | `robot-only; multi-view; action-labeled; scripted` | [site](https://sharinka0715.github.io/X-WAM/) | - | - |
| [Geometry-aware 4D Video Generation](https://arxiv.org/abs/2507.01099) | ICLR 2026 | R1 / R4 | Partial | `robot-only; multi-view; action-labeled; scripted` | [site](https://robot4dgen.github.io/) | [code](https://github.com/lzylucy/4dgen) | [ckpt](https://real.stanford.edu/4dgen/) |
| [Learning Action Manifold](https://arxiv.org/abs/2605.11832) | arXiv 2026 | Support / R4 | No | `robot-only; multi-view; action-labeled; scripted` | [site](https://junjxiao.github.io/Multi-view-VLA.github.io/) | [code](https://github.com/junjxiao/Multi-view-VLA) | - |
| [RoboFlow4D](https://arxiv.org/abs/2605.17522) | arXiv 2026 | R1 / R4 | Partial | `robot-only; exo; action-labeled; scripted` | [site](https://simonlinsx.github.io/RoboFlow4D_Page/) | - | - |
| [GaussianDream](https://arxiv.org/abs/2605.20752) | arXiv 2026 | Support / R4 | No | `robot-only; multi-view; action-labeled; scripted` | - | [code](https://github.com/TuojingAI/GaussianDream) | - |
| [GEM-4D](https://arxiv.org/abs/2605.22882) | arXiv 2026 | R1 / R4 | Partial | `robot-only; exo; action-labeled; scripted` | [site](https://anonymous-submission-20.github.io/gem.github.io/) | - | - |
| [DexWM](https://arxiv.org/abs/2512.13644) | arXiv 2025 | R4 | Partial | `co-training; ego; explicit-3D-labeled + action-labeled; unscripted` | [site](https://raktimgg.github.io/dexwm/) | [code](https://github.com/facebookresearch/dexwm) | - |
| [Donk](https://arxiv.org/abs/2606.03868) | arXiv 2026 | R2 / Explicit 3D | Yes | `dexterous; bimanual; explicit-3D-hand; action-labeled; scripted` | - | - | - |
| [WAM4D](https://arxiv.org/abs/2606.14048) | arXiv 2026 | R3 / Explicit 3D | Partial | `robot-only; explicit-4D; action-labeled; scripted` | - | - | - |
| [Structured 4D](https://arxiv.org/abs/2607.01166) | ICML 2026 | R4 / Explicit 3D | Partial | `robot-only; multi-view RGB-D; explicit-3D-latent; action-labeled; scripted` | [site](https://structured-4d-model.github.io/) | [code](https://github.com/sky-lzy/Structured-4D-Model) | [ckpt](https://huggingface.co/zhiyi24/structured-4d-model) |
| [PhysMani](https://arxiv.org/abs/2607.01938) | ECCV 2026 | R4 / Explicit 3D | Partial | `robot-only; multi-view RGB-D; explicit-3D-Gaussian; dynamic-object; action-labeled; scripted` | - | [code](https://github.com/vLAR-group/PhysMani) | - |
| [WSA1](https://arxiv.org/abs/2607.03941) | arXiv 2026 | R3 / Explicit 3D | Partial | `co-training; multi-view + ego; action-free + action-labeled; scripted` | - | [code](https://github.com/zaleni/WSA) | - |
| [RynnWorld-Teleop](https://arxiv.org/abs/2607.06558) | arXiv 2026 | Support / R4 | No | `co-training; ego; depth-aware-hand-pose; action-labeled; teleop-generated` | [site](https://alibaba-damo-academy.github.io/RynnWorld-Teleop.github.io/) | [code](https://github.com/alibaba-damo-academy/RynnWorld-Teleop) | [ckpt](https://huggingface.co/Alibaba-DAMO-Academy/RynnWorld-Teleop) / [causal](https://huggingface.co/Alibaba-DAMO-Academy/RynnWorld-Teleop-Causal) |
| [RynnWorld-4D](https://arxiv.org/abs/2607.06559) | arXiv 2026 | R4 / Explicit 3D | Partial | `co-training; ego; pseudo-RGB-DF; action-free + action-labeled; scripted` | [site](https://alibaba-damo-academy.github.io/RynnWorld-4D.github.io/) | [code](https://github.com/alibaba-damo-academy/RynnWorld-4D) | [ckpt](https://huggingface.co/Alibaba-DAMO-Academy/RynnWorld-4D) |
| [MECo-WAM](https://arxiv.org/abs/2607.05468) | arXiv 2026 | R3 / Explicit 3D | Partial | `robot-only; exo; action-labeled; scripted` | - | - | - |
| [GenVid2Robot](https://arxiv.org/abs/2607.09191) | arXiv 2026 | R1 / Explicit 3D | Partial | `robot-only; exo RGB-D; action-labeled; scripted` | - | - | - |

## 🧪 Physics / Functional Utility / Evaluation

| Paper | Venue | Route | Strict | Data tags | Project | Code | Checkpoint |
| --- | --- | --- | --- | --- | --- | --- | --- |
| [EWMBench](https://arxiv.org/abs/2505.09694) | arXiv 2025 | R4 | No | `robot-only; exo; action-labeled; scripted` | - | [code](https://github.com/AgibotTech/EWMBench) | - |
| [Genie Envisioner](https://arxiv.org/abs/2508.05635) | arXiv 2025 | Support / R4 | Partial | `robot-only; exo; action-labeled; scripted` | [site](https://genie-envisioner.github.io/) | [code](https://github.com/AgibotTech/Genie-Envisioner) | - |
| [WorldArena](https://arxiv.org/abs/2602.08971) | arXiv 2026 | R4 | No | `robot-only; exo; action-labeled; scripted` | [site](https://world-arena.ai/) | [code](https://github.com/tsinghua-fib-lab/WorldArena) | - |
| [WorldArena 2.0](https://arxiv.org/abs/2605.17912) | arXiv 2026 | R4 | No | `robot-only; visuotactile; action-labeled; scripted` | [site](https://world-arena.ai/) | [code](https://github.com/tsinghua-fib-lab/WorldArena) | - |
| [ContactGaussian-WM](https://arxiv.org/abs/2602.11021) | arXiv 2026 | R4 | No | `robot-only; multi-view; action-labeled; scripted` | [site](https://contactgaussian-wm.github.io/) | - | - |
| [Inference-time Physics Alignment](https://arxiv.org/abs/2601.10553) | arXiv 2026 | R3 / R4 | No | `robot-only; exo; action-labeled; scripted` | - | [code](https://github.com/facebookresearch/WMReward) | - |
| [WoVR](https://arxiv.org/abs/2602.13977) | arXiv 2026 | Support / R4 | No | `robot-only; exo; action-labeled; scripted` | - | - | - |
| [ReWorld](https://arxiv.org/abs/2601.12428) | arXiv 2026 | Support / R4 | No | `robot-only; exo; action-labeled; scripted` | - | - | - |
| [EVA](https://arxiv.org/abs/2603.17808) | arXiv 2026 | R1 / R4 | Partial | `robot-only; exo; action-labeled; scripted` | [site](https://eva-project-page.github.io/) | [code](https://github.com/RobbinW/EVA) | - |
| [World Action Verifier](https://arxiv.org/abs/2604.01985) | arXiv 2026 | R4 | No | `co-training; exo; action-free + action-labeled; scripted` | [site](https://world-action-verifier.github.io/) | [code](https://github.com/world-action-verifier/wav_minigrid) | - |
| [JailWAM](https://arxiv.org/abs/2604.05498) | arXiv 2026 | R4 | No | `robot-only; exo; action-labeled; scripted` | [site](https://jailwam.github.io/) | - | - |
| [ViVa](https://arxiv.org/abs/2604.08168) | arXiv 2026 | R4 | No | `robot-only; exo; action-labeled; scripted` | [site](https://viva-value-model.github.io/) | [code](https://github.com/GigaAI-research/ViVa) | - |
| [RoboWM-Bench](https://arxiv.org/abs/2604.19092) | arXiv 2026 | R4 | No | `robot-only; exo; action-labeled; scripted` | - | [code](https://github.com/fffstrong/RoboWM-Bench) | - |
| [Hi-WM](https://arxiv.org/abs/2604.21741) | arXiv 2026 | Support / R4 | No | `robot-only; exo; action-labeled; scripted` | [site](https://hi-wm.github.io/) | - | - |
| [dWorldEval](https://arxiv.org/abs/2604.22152) | arXiv 2026 | R4 | No | `robot-only; exo; action-labeled; scripted` | [site](https://dworldeval.github.io/) | - | - |
| [EA-WM](https://arxiv.org/abs/2605.06192) | arXiv 2026 | R4 | No | `robot-only; exo; action-labeled; scripted` | - | - | - |
| [When to Trust Imagination](https://arxiv.org/abs/2605.06222) | arXiv 2026 | Support / R4 | No | `robot-only; exo; action-labeled; scripted` | - | - | - |
| [Is the Future Compatible?](https://arxiv.org/abs/2605.07514) | arXiv 2026 Technical Report | R4 | No | `robot-only; exo; action-labeled; scripted` | - | - | - |
| [RoboAlign-R1](https://arxiv.org/abs/2605.03821) | arXiv 2026 | R4 | No | `robot-only; exo; action-labeled; scripted` | - | [code](https://github.com/Alexander-wu/RoboAlign_R1) | - |
| [DreamAvoid](https://arxiv.org/abs/2605.11750) | arXiv 2026 | Support / R4 | No | `robot-only; multi-view; action-labeled; scripted` | - | [code](https://github.com/XianzheFan/DreamAvoid) | - |
| [Feedback World Model](https://arxiv.org/abs/2605.15705) | arXiv 2026 | Support / R4 | No | `robot-only; exo; action-labeled; scripted` | [site](https://lorenzo-0-0.github.io/Feedback_World_Model/) | - | - |
| [Beyond Euclidean Proximity](https://arxiv.org/abs/2605.22164) | arXiv 2026 | R4 | No | `robot-only; exo; action-labeled; scripted` | - | - | - |
| [GE-Sim 2.0](https://arxiv.org/abs/2605.27491) | arXiv 2026 | Support / R4 | Partial | `robot-only; multi-view; action-labeled; scripted` | [site](https://ge-sim-v2.github.io/) | [code](https://github.com/AgibotTech/GE-Sim-V2) | - |
| [FAWAM](https://arxiv.org/abs/2606.08555) | arXiv 2026 | R2 / R4 | Partial | `robot-only; visuotactile; action-labeled; scripted` | - | - | - |
| [SKIP](https://arxiv.org/abs/2606.00664) | arXiv 2026 | Support / R4 | No | `robot-only; exo; action-labeled; scripted` | - | - | - |
| [Beyond Task Success](https://arxiv.org/abs/2606.01095) | arXiv 2026 | R4 | No | `evaluation; WAM + VLA diagnostics` | - | - | - |
| [RoboDream](https://arxiv.org/abs/2606.02577) | arXiv 2026 | Support / R4 | No | `robot-only; exo; action-labeled; scripted` | [site](https://junjieye.com/RoboDream/) | [code](https://github.com/Jay-Ye/RoboDream) | - |
| [PiL-World](https://arxiv.org/abs/2606.05773) | arXiv 2026 | R4 | No | `robot-only; exo; action-labeled; scripted` | - | - | - |
| [PRISM](https://arxiv.org/abs/2606.07974) | arXiv 2026 | Support / R4 | No | `robot-only; exo; action-labeled; scripted` | [site](https://yuhaiw.github.io/PRISM_web/) | - | - |
| [Video2Sim2Real](https://arxiv.org/abs/2606.08828) | arXiv 2026 | Support / R4 | No | `human-video; dexterous; action-free-to-robot; scripted` | [site](https://video2sim2real.github.io/) | - | - |
| [Targeting World Models](https://arxiv.org/abs/2606.09499) | arXiv 2026 | R4 Safety | No | `safety; robot world-model pipeline` | - | - | - |
| [EA-WM](https://arxiv.org/abs/2606.13053) | arXiv 2026 | R4 | No | `robot-only; exo; action-labeled; scripted` | - | - | - |
| [WEAVER](https://arxiv.org/abs/2606.13672) | arXiv 2026 | R4 / learned simulator | Partial | `robot-only; exo; action-labeled; scripted` | [site](https://arnavkj1995.github.io/WEAVER/) | [code](https://github.com/arnavkj1995/WEAVER) | [ckpt](https://huggingface.co/arnavkj1995/WEAVER) |
| [Dream-Tac](https://arxiv.org/abs/2606.08737) | arXiv 2026 | R2 / R4 | Partial | `robot-only; visuotactile; action-labeled; scripted` | - | [code](https://github.com/LYFCLOUDFAN/Dream-Tac) | - |
| [Tactile-WAM](https://arxiv.org/abs/2606.26663) | arXiv 2026 | R2 / R4 | Partial | `robot-only; visuotactile; action-labeled; scripted` | - | - | - |
| [WAM-RL](https://arxiv.org/abs/2606.17906) | CVPR Workshops 2026 | Support / R4 | No | `robot-only; exo; action-labeled + online RL; scripted` | - | - | - |
| [SC3-Eval](https://arxiv.org/abs/2606.18610) | arXiv 2026 | R4 | No | `evaluation; multi-view; action-conditioned world-model` | [site](https://weichengtseng.github.io/sc3-eval/) | - | - |
| [SafeDojo](https://arxiv.org/abs/2606.20698) | arXiv 2026 | R4 Safety / VAM-RL | No | `safe-RL; robot-only; exo; action-labeled; scripted` | - | - | - |
| [World Value Models](https://arxiv.org/abs/2606.24742) | arXiv 2026 | R4 / VAM-RL | No | `robot-only; exo; mixed-quality trajectories; scripted` | [site](https://zh1hao.wang/wvm/) | [code](https://github.com/255isWhite/WVM) | - |
| [REGEN](https://arxiv.org/abs/2606.27374) | arXiv 2026 | Support / R4 | No | `robot-only; exo; action-labeled; continual-learning` | [site](https://manishgovind.github.io/REGEN/) | [code](https://github.com/ManishGovind/REGEN) | - |
| [PhysisForcing](https://arxiv.org/abs/2606.28128) | arXiv 2026 | Support / R4 | No | `robot-only; exo; action-conditioned video; scripted` | [site](https://dagroup-pku.github.io/PhysisForcing.github.io/) | [code](https://github.com/DAGroup-PKU/PhysisForcing) | - |
| [RoboWorld](https://arxiv.org/abs/2607.01060) | arXiv 2026 | R4 | No | `evaluation; multi-view; action-labeled; real-policy rollouts` | - | - | - |
| [WorldSample](https://arxiv.org/abs/2607.02431) | arXiv 2026 | R4 / VAM-RL | No | `robot-only; real-robot; action-labeled + online RL; mixed real/synthetic` | [site](https://xxreinsno.github.io/worldsample/) | - | - |
| [DynaWM](https://arxiv.org/abs/2607.02604) | arXiv 2026 | Support / R4 | No | `robot-only; multi-view; action-labeled; simulated dynamic manipulation` | - | - | - |
| [GigaWorld-1](https://arxiv.org/abs/2607.02642) | arXiv 2026 | R4 | No | `mixed robot + human-ego; multi-view; action-labeled; policy rollouts` | [site](https://open-gigaai.github.io/giga-world-1/) | [code](https://github.com/open-gigaai/giga-world-1) | - |
| [HALO-WA](https://arxiv.org/abs/2607.04265) | arXiv 2026 | R3 / VAM-RL | No | `robot-only; multi-view; action-labeled + online RL; scripted` | - | [code](https://github.com/YeanRoot/HALO-WA) | - |
| [WAM-TTT](https://arxiv.org/abs/2607.06988) | arXiv 2026 | Support / R4 | Partial | `human-video + robot; action-free human video + action-labeled robot; real household` | - | - | - |

<a id="vam-rl-view"></a>
## 🕹️ VAM-RL View

| Paper | Base route reading | Optimization mechanism | Why it gets the VAM-RL tag |
| --- | --- | --- | --- |
| [VAMPO](https://arxiv.org/abs/2603.19370) | `R1` | `GRPO` post-training of a VPP-style video predictor with Euler Hybrid sampling | the staged predictive-latent actor remains intact, so the novelty is RL post-training that sharpens control-relevant visual dynamics |
| [NoiseGate](https://arxiv.org/abs/2605.07794) | `R2` | `GRPO` trains a lightweight gate-policy network over per-latent denoising timesteps | the joint WAM backbone remains intact, while task success learns how much imagined future each action-denoising step should trust |
| [SANTS](https://arxiv.org/abs/2605.27947) | `Support / R2` | PPO-style training of a state-adaptive scheduler over video denoising depth and jump size | the WAM backbone remains frozen while RL learns which intermediate future state is most action-useful, making inference scheduling a control objective rather than a generic sampler shortcut |
| [EVA](https://arxiv.org/abs/2603.17808) | `R1 / R4` | `Flow-GRPO` with IDM-derived executability rewards over staged video planning rollouts | the paper keeps video-world-model planning and IDM action recovery intact, but optimizes the generator toward more executable futures |
| [WoVR](https://arxiv.org/abs/2602.13977) | `Support / R4` | `masked GRPO` inside a controllable video world model with Keyframe-Initialized Rollouts and low-frequency policy-world co-evolution | the backbone remains a world-model-based simulator for VLA post-training, while the novelty is reliability-aware RL that explicitly regulates hallucination and distribution shift |
| [ReWorld](https://arxiv.org/abs/2601.12428) | `Support / R4` | PPO-style post-training of embodied world models with a multi-dimensional preference reward model | the backbone remains a video-based embodied world model, while the headline contribution is reward modeling plus RL alignment toward physical realism, task logic, and embodiment plausibility |
| [AIM](https://arxiv.org/abs/2604.11135) | `R2` | self-distillation RL with dense ASVM reward and sparse task-success reward while freezing the video/value-map branches | the backbone remains a joint video-action model with an explicit spatial value-map bridge, while RL sharpens the action head to follow the future-grounded value map in closed-loop control |
| [ViVa](https://arxiv.org/abs/2604.08168) | `R4` | video-generative value modeling for robot RL, with future dynamics used as the value-estimation substrate | the value model is trained to anticipate embodied futures before scoring behavior, so the optimization signal becomes world-model-derived rather than a static state score |
| [RoboAlign-R1](https://arxiv.org/abs/2605.03821) | `R4` | `GRPO` with a distilled multimodal robot-video reward and sliding-window re-encoding | the video world model is optimized toward instruction following, manipulation success, contact, and physics signals instead of reconstruction-only metrics |
| [Hi-WM](https://arxiv.org/abs/2604.21741) | `Support / R4` | human corrective post-training inside an action-conditioned world model with rollback, branching, and later real-world consolidation | the world model becomes the post-training substrate that turns human interventions into reusable imagined correction data rather than one-off offline labels |
| [WAM-RL](https://arxiv.org/abs/2606.17906) | `Support / R4` | Flow-SDE actor RL with reconstruction-based dense rewards plus online video SFT of the world model with KL regularization | the WA backbone remains a world-model-plus-actor system, while online RL co-optimizes imagined futures and executable actions from environment feedback |
| [SafeDojo](https://arxiv.org/abs/2606.20698) | `Support / R4` | safety-constrained `GRPO` over an interactive video world model, using imagined task reward and safety cost with Lagrangian balancing | the world model supplies imagined rollouts and reward/cost estimates, so policy improvement is explicitly RL-based and safety-constrained rather than only supervised VLA fine-tuning |
| [MV-WAM](https://arxiv.org/abs/2606.21088) | `R2 / R4` | manifold-aware multi-objective training with Monte-Carlo progress-value tokens and value-guided rollback during online execution | the model jointly predicts video, actions, and value; the value signal monitors progress/misalignment and actively regulates recovery, making it a value-optimization boundary case for the VAM-RL view |
| [World Value Models](https://arxiv.org/abs/2606.24742) | `R4` | world-model-backed value-flow training with video/value DiT MoT coupling, value prefix randomization, and downstream use in AWR / Filtered BC | it repurposes temporal/future-planning priors from video world models into a generalist value signal that improves robot policy learning from mixed-quality data |
| [WorldSample](https://arxiv.org/abs/2607.02431) | `Support / R4` | closed-loop real-robot RL with world-model synthetic replay and Policy-Paced Learning | the world model generates counterfactual replay around real rollouts, while policy-paced scheduling controls when synthetic experience is trusted during real-robot RL |
| [HALO-WA](https://arxiv.org/abs/2607.04265) | `R3 / Support` | online actor-critic adapter trained over frozen WA reference actions and visual latents | the WA backbone stays frozen, so the optimization target is a lightweight latent-guided adapter that turns WAM foresight into fast precision adaptation |
| [WAM-TTT](https://arxiv.org/abs/2607.06988) | `Support / R4` | test-time training writes human demonstration videos into frozen-WAM fast-weight memory | it is not classic RL, but it belongs as a VAM-RL boundary case because deployment-time optimization changes how the frozen WAM conditions action generation from raw human video |

Awesome-Video-Action-Model

Explore Similar Repositories

Was this content helpful?

Recent Feedback

Download README