IffYuan/Embodied-R1-3B-v1

Embodied-R1-3B-v1

Embodied-R1: Reinforced Embodied Reasoning for General Robotic Manipulation (ICLR 2026)

[🌐 Project Website] [📄 Paper] [🏆 ICLR2026 Version] [🎯 Dataset] [📦 Code]

Model Details

Model Description

Embodied-R1 is a 3B vision-language model (VLM) for general robotic manipulation. It introduces a Pointing mechanism and uses Reinforced Fine-tuning (RFT) to bridge perception and action, with strong zero-shot generalization in embodied tasks.

Figure: Embodied-R1 framework, performance overview, and zero-shot manipulation demos.

Model Sources

• —Repository: https://github.com/pickxiguapi/Embodied-R1
• —Paper: http://arxiv.org/abs/2508.13998
• —OpenReview: https://openreview.net/forum?id=i5wlozMFsQ

Updates

• —[2026-03] VABench-P / VABench-V released: VABench-P, VABench-V
• —[2026-03-03] Embodied-R1 dataset released: https://huggingface.co/datasets/IffYuan/Embodied-R1-Dataset
• —[2026-01-27] Accepted by ICLR 2026
• —[2025-08-22] Embodied-R1-3B-v1 checkpoint released

Intended Uses

Direct Use

This model is intended for research and benchmarking in embodied reasoning and robotic manipulation tasks, including:

• —Visual target grounding (VTG)
• —Referring region grounding (RRG/REG-style tasks)
• —Open-form grounding (OFG)

Out-of-Scope Use

• —Safety-critical real-world deployment without additional safeguards and validation
• —Decision-making in high-risk domains
• —Any use requiring guaranteed robustness under distribution shift

How to Use

Setup

git clone https://github.com/pickxiguapi/Embodied-R1.git
cd Embodied-R1

conda create -n embodied_r1 python=3.11 -y
conda activate embodied_r1

pip install transformers==4.51.3 accelerate
pip install qwen-vl-utils[decord]

Inference

python inference_example.py

Example Tasks

• —VTG: put the red block on top of the yellow block
• —RRG: put pepper in pan
• —REG: bring me the camel model
• —OFG: loosening stuck bolts

(Visualization examples are available in the project repo: assets/)

Evaluation

cd eval
python hf_inference_where2place.py
python hf_inference_vabench_point.py
...

Related benchmarks:

• —Embodied-R1-Dataset
• —VABench-P
• —VABench-V

Training

Training scripts are available at: https://github.com/pickxiguapi/Embodied-R1/tree/main/scripts

# Stage 1 training
bash scripts/stage_1_embodied_r1.sh

# Stage 2 training
bash scripts/stage_2_embodied_r1.sh

Key files:

• —scripts/config_stage1.yaml
• —scripts/config_stage2.yaml
• —scripts/stage_1_embodied_r1.sh
• —scripts/stage_2_embodied_r1.sh
• —scripts/model_merger.py (checkpoint merging + HF export)

Limitations

• —Performance may vary across environments, camera viewpoints, and unseen object domains.
• —Outputs are generated from visual-language reasoning and may include localization/action errors.
• —Additional system-level constraints (calibration, motion planning, safety checks) are required for real robot deployment.

Citation

@article{yuan2026embodied,
  title={Embodied-r1: Reinforced embodied reasoning for general robotic manipulation},
  author={Yuan, Yifu and Cui, Haiqin and Huang, Yaoting and Chen, Yibin and Ni, Fei and Dong, Zibin and Li, Pengyi and Zheng, Yan and Tang, Hongyao and Hao, Jianye},
  journal={The Fourteenth International Conference on Learning Representations},
  year={2026}
}

@article{yuan2026seeing,
  title={From seeing to doing: Bridging reasoning and decision for robotic manipulation},
  author={Yuan, Yifu and Cui, Haiqin and Chen, Yibin and Dong, Zibin and Ni, Fei and Kou, Longxin and Liu, Jinyi and Li, Pengyi and Zheng, Yan and Hao, Jianye},
  journal={The Fourteenth International Conference on Learning Representations},
  year={2026}
}

Acknowledgements

If this model or resources are useful for your research, please consider citing our work and starring the repository.