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

ABot-AgentOS

★ 36 Open GitHub ↗

A General Robotic Agent OS with Lifelong Multi-modal Memory

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# ABot-AgentOS

### A General Robotic Agent OS with Lifelong Multi-modal Memory

AMAP CV Lab

---

## 📰 News

- **2026-07-11** — We released the [ABot-AgentOS technical report](https://arxiv.org/abs/2607.10350).
- **2026-07-21** — We released the [ABot-AgentOS project page](https://amap-cvlab.github.io/ABot-AgentOS).

## 🔥 Overview

**ABot-AgentOS** is a general Agent OS for embodied robotic agents. It
sits between foundation models and robot control stacks, coordinating
multi-modal perception, LLM/VLM reasoning, context management, skill
execution, tool use, verification, memory, and training for deployable
robot agents.

Rather than treating a robot agent as a single monolithic model,
ABot-AgentOS provides a modular runtime and training framework for
long-horizon embodied tasks across heterogeneous robot embodiments.

## ✨ Highlights

- **Agent OS architecture**: edge-cloud LLM routing, Agent Harness,
  skills/tools, memory, and robot hardware interfaces in one runtime.
- **Agent Harness**: a reasoning-execution-verification loop with a
  main LLM, Skill Runner, Verifier, Context Management, and Skill
  Evolvement.
- **EmbodiedWorldBench**: an executable long-horizon embodied benchmark
  spanning indoor, outdoor, and hybrid scenarios.
- **Memory + self-evolution**: typed, source-grounded multi-modal graph
  memory with traceable retrieval and split-wise failure-driven
  evo-assets.
- **Training pipeline**: text-based embodied environments, teacher
  trajectories, SFT/RL, and a self-evolving LLM-as-a-Judge reward
  engine for smaller deployable models.

## 📊 Evaluation

Initial evaluation covers both embodied execution and memory:

- On the EmbodiedWorldBench subset, ABot-AgentOS improves TSR/GCR over
  a single-controller ReAct baseline under the same main model.
- On representative long-term memory and embodied QA benchmarks,
  ABot-AgentOS graph memory and **+ Self-evo** show gains over previous
  references shown in the report.

Please see the [technical report](https://arxiv.org/abs/2607.10350) and
[project page](https://amap-cvlab.github.io/ABot-AgentOS) for details.

## 🚧 Release Status

The technical report and project page are available. Code and resources
are being prepared for open-source release.

## 📝 Citation

If you find this work useful, please consider citing our paper.

```bibtex
@article{tian2026abot,
  title={ABot-AgentOS: A General Robotic Agent OS with Lifelong Multi-modal Memory},
  author={Tian, Jiayi and Liu, Shiao and Xu, Yuting and Lu, Jia and Guan, Zihao and Han, Honglin and Yang, Di and Gu, Minqi and Qian, Yifei and Zhang, Tianlin and others},
  journal={arXiv preprint arXiv:2607.10350},
  year={2026}
}
```

## 📄 License

This project is released under the MIT License.

## 🙏 Acknowledgments

This work is developed by AMAP CV Lab.

ABot-AgentOS

Explore Similar Repositories

Was this content helpful?

Recent Feedback

Download README