Spatial-SSRL - GitLite

KO | EN

gitlite — search

#javascript #python #react #ai #android #machine-learning #deep-learning #macos #nodejs #java #kubernetes #vue #openai #open-source

Spatial-SSRL

(★ 136)

[CVPR 2026] Official release of "Spatial-SSRL: Enhancing Spatial Understanding via Self-Supervised Reinforcement Learning"

File Explorer

Download Latest Version (.zip)

- comparison_v2.png
- exp_result.png
- pipeline_1029final.png
- teaser_1029final.png
- - RGB-D_depth1.h5
  - RGB-D_depth2.h5
  - RGB-D_depth3.npy
  - RGB-D_depth_mask3.npy
  - RGB-D_img1.jpg
  - RGB-D_img2.jpg
  - RGB-D_img3.png
  - RGB_img.png
- Crop.py
- Depth.py
- Flip.py
- Position.py
- Shuffle.py
- - .env
  - run.py
- - main.py
  - run_cot.sh
  - run_nocot.sh
- - check_accuracy.py
  - run_cot.sh
  - run_nocot.sh
  - vsi_infer.py
- - main.py
  - model_config.json
  - run_cot.sh
  - run_nocot.sh
- Eval.md
- requirements.txt
- case1.jpg
- case2.jpg
- eg1.jpg
- eg2.jpg
- eg3.jpg
- eg4.jpg
- eg5.jpg
- eg6.jpg
- eg_qwen3vl.jpg
- record.txt
- - .gitignore
  - .pre-commit-config.yaml
  - Dockerfile
  - Dockerfile.legacy
  - LICENSE
  - Makefile
  - pyproject.toml
  - README.md
  - requirements.txt
  - setup.py
- - .dockerignore
  - .env.local
  - .gitattributes
  - .gitignore
  - .pre-commit-config.yaml
  - CITATION.cff
  - ds_z3_config.json
  - LICENSE
  - Makefile
  - MANIFEST.in
  - pyproject.toml
  - README.md
  - README_zh.md
  - requirements.txt
  - setup.py
- training.md
demo.py
LICENSE
README.md

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

<div align="center">
  <h1 align="center">
    <div style="display: flex; align-items: center; justify-content: center;">
      <div style="text-align: left; line-height: 1.3;">
        (CVPR 2026)Spatial-SSRL: Enhancing Spatial Understanding via Self-Supervised Reinforcement Learning
      </div>
    </div>
  </h1>
  <p align="center">
    <a href="https://scholar.google.com/citations?hl=en&user=yQ8U3tUAAAAJ"><strong>Yuhong Liu</strong></a >
    ·
    <a href="https://beichenzbc.github.io/"><strong>Beichen Zhang</strong></a >
    ·
    <a href="https://yuhangzang.github.io/"><strong>Yuhang Zang<sup>&dagger;</sup></strong></a >
    ·
    <a href="https://scholar.google.com/citations?user=sJkqsqkAAAAJ"><strong>Yuhang Cao</strong></a >
    ·
    <a href="https://github.com/Cooperx521"><strong>Long Xing</strong></a >
    </br>
    <a href="https://lightdxy.github.io/"><strong>Xiaoyi Dong</strong></a >
    ·
    <a href="https://github.com/kennymckormick"><strong>Haodong Duan</strong></a >
    ·
    <a href="http://dahua.site/"><strong>Dahua Lin</strong></a >
    ·
     <a href="https://myownskyw7.github.io/"><strong>Jiaqi Wang<sup>&dagger;</sup></strong></a >
  </p >
  <p align="center" style="font-size: 1em; margin-top: -1em">  <sup>&dagger;</sup>Corresponding authors. </p >
  <p align="center" style="font-size: 1.2em; margin-top: 0.5em">
    📖<a href="https://arxiv.org/abs/2510.27606">Paper</a>
  | 🏠<a href="https://github.com/InternLM/Spatial-SSRL">Homepage</a >
  | 🤗<a href="https://huggingface.co/internlm/Spatial-SSRL-7B">Spatial-SSRL-7B Model</a >
  | 🤗<a href="https://huggingface.co/internlm/Spatial-SSRL-3B">Spatial-SSRL-3B Model</a >
  | 🤗<a href="https://huggingface.co/internlm/Spatial-SSRL-Qwen3VL-4B">Spatial-SSRL-Qwen3VL-4B Model</a >
  | 🤗<a href="https://huggingface.co/datasets/internlm/Spatial-SSRL-81k">Spatial-SSRL-81k Dataset</a >
  | 📰<a href="https://huggingface.co/papers/2510.27606">Daily Paper</a >
  </p > 
</div>

## 📢 News
- 🚀 [2026/04/05] We have released the training code of Spatial-SSRL.
- 🚀 [2026/02/25] We have released the [🤗Spatial-SSRL-3B Model](https://huggingface.co/internlm/Spatial-SSRL-3B), initialized from Qwen2.5-VL-3B-Instruct.
- 🚀 [2026/02/21] Our work has been accepted by CVPR 2026.
- 🚀 [2025/11/24] We have released the [🤗Spatial-SSRL-Qwen3VL-4B Model](https://huggingface.co/internlm/Spatial-SSRL-Qwen3VL-4B), initialized from Qwen3-VL-4B-Instruct.
- 🚀 [2025/11/03] Now you can try out Spatial-SSRL-7B on [🤗Spatial-SSRL Space](https://huggingface.co/spaces/yuhangzang/Spatial-SSRL).
- 🚀 [2025/11/03] We have released the [Spatial-SSRL-7B Model](https://huggingface.co/internlm/Spatial-SSRL-7B), and [Spatial-SSRL-81k Dataset](https://huggingface.co/datasets/internlm/Spatial-SSRL-81k).
- 🚀 [2025/11/02] We have released the Spatial-SSRL [repository](https://github.com/InternLM/Spatial-SSRL).

## 🌈 Overview
We are thrilled to introduce <strong>Spatial-SSRL</strong>, a novel self-supervised RL paradigm aimed at enhancing LVLM spatial understanding. 
By optimizing Qwen2.5-VL-7B with Spatial-SSRL, the model exhibits stronger spatial intelligence across seven spatial understanding benchmarks in both image and video settings.
</p>
<p style="text-align: center;"> 
  <img src="assets/teaser_1029final.png" alt="Teaser" width="100%"> 
</p>
Spatial-SSRL is a <strong>lightweight</strong> tool-free framework that is natually compatible with the RLVR training paradigm and easy to extend to a multitude of pretext tasks.
Five tasks are currently formulated in the framework, requiring only ordinary RGB and RGB-D images. <strong>And we welcome you to join Spatial-SSRL with effective pretext tasks to further strengthen the capabilities of LVLMs!</strong>

<p style="text-align: center;"> 
  <img src="assets/pipeline_1029final.png" alt="Pipeline" width="100%"> 
</p>

## 💡 Highlights
- 🔥 **Highly Scalable:** Spatial-SSRL uses ordinary raw RGB and RGB-D images instead of richly-annotated public datasets or manual labels for data curation, making it highly scalable.
- 🔥 **Cost-effective:** Avoiding the need for human labels or API calls for general LVLMs throughout the entire pipeline endows Spatial-SSRL with cost-effectiveness.
- 🔥 **Lightweight:** Prior approaches for spatial understanding heavily rely on annotation of external tools, incurring inherent errors in training data and additional cost. In constrast, Spatial-SSRL is completely tool-free and can easily be extended to more self-supervised tasks. 
- 🔥 **Naturally Verifiable:** Intrinsic supervisory signals determined by pretext objectives are naturally verifiable, aligning Spatial-SSRL well with the RLVR paradigm.
<p style="text-align: center;"> 
  <img src="assets/comparison_v2.png" alt="Teaser" width="100%"> 
</p>

## 📊 Results
We train Qwen2.5-VL-3B and Qwen2.5-VL-7B with our Spatial-SSRL paradigm and the experimental results across seven spatial understanding benchmarks are shown below.
<p style="text-align: center;"> 
  <img src="assets/exp_result.png" alt="Pipeline" width="100%"> 
</p>

## ⭐️ Quick Start
To directly experience <strong>Spatial-SSRL-7B</strong>, you can try it out on [🤗Spatial-SSRL Space](https://huggingface.co/spaces/yuhangzang/Spatial-SSRL)!

Here we provide a code snippet for you to start a simple trial of <strong>Spatial-SSRL-7B</strong> on your own device. You can download the model from 🤗<a href="https://huggingface.co/internlm/Spatial-SSRL-7B">Spatial-SSRL-7B Model</a > before your trial!
</p>

```python
from transformers import Qwen2_5_VLForConditionalGeneration, AutoTokenizer, AutoProcessor
from qwen_vl_utils import process_vision_info

model_path = "internlm/Spatial-SSRL-7B" #You can change it to your own local path if deployed already
img_path = "examples/eg1.jpg"
question = "Consider the real-world 3D locations of the objects. Which object has a higher location? A. yellow bear kite B. building"
#We recommend using the format prompt to make the inference consistent with training
format_prompt = "\n You FIRST think about the reasoning process as an internal monologue and then provide the final answer. The reasoning process MUST BE enclosed within <think> </think> tags. The final answer MUST BE put in \\boxed{}."

model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
    model_path, torch_dtype="auto", device_map="auto"
)
processor = AutoProcessor.from_pretrained(model_path)
messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "image",
                "image": img_path,
            },
            {"type": "text", "text": question + format_prompt},
        ],
    }
]

text = processor.apply_chat_template(
    messages, tokenize=False, add_generation_prompt=True
)
image_inputs, video_inputs = process_vision_info(messages)
inputs = processor(
    text=[text],
    images=image_inputs,
    videos=video_inputs,
    padding=True,
    return_tensors="pt",
)
inputs = inputs.to("cuda")

generated_ids = model.generate(**inputs, max_new_tokens=4096, do_sample=False)
generated_ids_trimmed = [
    out_ids[len(in_ids) :] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
]
output_text = processor.batch_decode(
    generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
)
print("Model Response:", output_text)
```

Here we provide a code snippet for you to start a simple trial of <strong>Spatial-SSRL-Qwen3VL-4B</strong> on your own device. You can download the model from 🤗<a href="https://huggingface.co/internlm/Spatial-SSRL-Qwen3VL-4B">Spatial-SSRL-Qwen3VL-4B Model</a > before your trial!
</p>

```python
from transformers import AutoProcessor, AutoModelForImageTextToText #transformers==4.57.1
from qwen_vl_utils import process_vision_info #0.0.14
import torch

model_path = "internlm/Spatial-SSRL-Qwen3VL-4B" #You can change it to your own local path if deployed already

#Change the path of the input image
img_path = "examples/eg_qwen3vl.jpg"

#Change your question here
question = "Question: Consider the real-world 3D locations and orientations of the objects. If I stand at the man's position facing where it is facing, is the menu on the left or right of me?\nOptions:\nA. on the left\nB. on the right\n"

question += "Please select the correct answer from the options above. \n"
#We recommend using the format prompt to make the inference consistent with training
format_prompt = "You FIRST think about the reasoning process as an internal monologue and then provide the final answer. The reasoning process MUST BE enclosed within <think> </think> tags. The final answer MUST BE put in \\boxed{}."

model = AutoModelForImageTextToText.from_pretrained(
    model_path, torch_dtype=torch.float16, device_map='auto', attn_implementation='flash_attention_2'
    )
processor = AutoProcessor.from_pretrained(model_path)

messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "image",
                "image": img_path,
            },
            {"type": "text", "text": question + format_prompt},
        ],
    }
]

text = processor.apply_chat_template(
    messages, tokenize=False, add_generation_prompt=True
)
image_inputs, video_inputs = process_vision_info(messages)
inputs = processor(
    text=[text],
    images=image_inputs,
    videos=video_inputs,
    padding=True,
    return_tensors="pt",
)
inputs = inputs.to("cuda")

generated_ids = model.generate(**inputs, max_new_tokens=4096, do_sample=False)
generated_ids_trimmed = [
    out_ids[len(in_ids) :] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
]
output_text = processor.batch_decode(
    generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
)
print("Model Response:", output_text[0])
```
## 🛠️ Training
See <a href="https://github.com/InternLM/Spatial-SSRL/blob/main/training/training.md">training.md</a >

## 🛠️ Evaluation
Prepare your environment:
```bash
git clone https://github.com/InternLM/Spatial-SSRL.git
conda create -n spatialssrl python==3.10
conda activate spatialssrl
cd Spatial-SSRL/evaluation
pip install -r requirements.txt

# Recommended
pip install flash-attn --no-build-isolation
```
Start your evaluation by referring to the tutorials in <a href="https://github.com/InternLM/Spatial-SSRL/blob/main/evaluation/Eval.md">Eval.md</a >

## Cases
<p style="text-align: center;"> 
  <img src="examples/case1.jpg" alt="Teaser" width="100%"> 
</p>
<p style="text-align: center;"> 
  <img src="examples/case2.jpg" alt="Teaser" width="100%"> 
</p>

## ✒️Citation
If you find this project useful, please kindly cite:
```
@article{liu2025spatial,
  title={Spatial-SSRL: Enhancing Spatial Understanding via Self-Supervised Reinforcement Learning},
  author={Liu, Yuhong and Zhang, Beichen and Zang, Yuhang and Cao, Yuhang and Xing, Long and Dong, Xiaoyi and Duan, Haodong and Lin, Dahua and Wang, Jiaqi},
  journal={arXiv preprint arXiv:2510.27606},
  year={2025}
}
```

## 📄 License
![Code License](https://img.shields.io/badge/Code%20License-Apache_2.0-green.svg) ![Data License](https://img.shields.io/badge/Data%20License-CC%20By%20NC%204.0-red.svg)

**Usage and License Notices**: The data and code are intended and licensed for research use only.

## ❤️ Acknowledgement
We extend our sincere gratitude to <a href="https://github.com/open-compass/VLMEvalKit">VLMEvalkit</a >, the powerful toolkit to evaluate a vast range of LMMs!

The work is built upon <a href="https://github.com/hiyouga/EasyR1">EasyR1</a > and <a href="https://github.com/hiyouga/LLaMAFactory">LlaMaFactory</a >, two excellent codebase for LLM training!