Embodied-R1: Reinforced Embodied Reasoning for General Robotic Manipulation (ICLR2026)

---

🔥 Updates

• [2026-03] 🤗 VABench-P and VABench-V benchmarks are now available on Hugging Face Datasets: [VABench-P] [VABench-V]

• [2026-03-03] 📦 Dataset Open-Sourced! The Embodied-R1 dataset is now publicly available at Hugging Face Datasets.

• [2026-01-27] 🏆 Accepted by (ICLR2026)! Embodied-R1 are accepted to The Fourteenth International Conference on Learning Representations (ICLR2026).

• [2025-08-22] 🤗 Model Released! The Embodied-R1 3B v1 checkpoint is now available at Hugging Face Model Hub.

• [2025-08-21] 🚀 Inference Scripts Released! We have released our inference prompts and scripts for embodied pointing abilities.

• [2025-08-20] 📚 Models and Datasets Released! We have released our pre-trained models, training datasets, and comprehensive evaluation benchmarks. Check out our HuggingFace collection for all available resources.

---

📖 Overview

Embodied-R1 is a 3B vision-language model (VLM) designed for general robotic manipulation. Through an innovative "Pointing" mechanism and Reinforced Fine-tuning (RFT) training methodology, it effectively bridges the "seeing-to-doing" gap in robotics, achieving remarkable zero-shot generalization capabilities.

Figure 1: Embodied-R1 framework overview, comprehensive performance evaluation, and zero-shot robotic manipulation demonstrations.

---

🛠️ Setup

1. Clone the repository:

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

2. Create and activate Conda environment:

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

3. Install dependencies for inference:

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

4. Install dependencies for training (optional):

   pip install -r requirements.txt
   

---

🚀 Inference

Run the example code:

cd Embodied-R1/
python inference_example.py

VTG Example

Task instruction: put the red block on top of the yellow block

Before prediction (original image):

After prediction (visualization result):

RRG Example

Task instruction: put pepper in pan

Before prediction (original image):

After prediction (visualization result):

REG Example

Task instruction: bring me the camel model

Before prediction (original image):

After prediction (visualization result):

OFG Example

Task instruction: loosening stuck bolts

Before prediction (original image):

After prediction (visualization result):

---

📊 Evaluation

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

🧠 Training

Training scripts are available in scripts/:

# Stage 1 training
bash scripts/stage_1_embodied_r1.sh

# Stage 2 training (set your stage-1 checkpoint path first)
bash scripts/stage_2_embodied_r1.sh

Key training 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 (for checkpoint merging and HF export)

📜 Citation

If you use our work in your research, please cite:

@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}
}