Diffusion-Based Imaginative Coordination for Bimanual Manipulation

Figure 1: Framework overview of our diffusion-based policy.

Figure 2: Task visualization and results overview (2 ALOHA + 16 RoboTwin + 4 Real-world tasks).

📰 News

June 25th, 2025: Our paper is accepted by ICCV 2025.

May 20th, 2025: We released our code and model.

Clone the source code

https://github.com/return-sleep/Diffusion_based_imaginative_Coordination.git
cd Diffusion_based_imaginative_Coordination

ALOHA

🔧 Installation

Install the required packages, see INSTALLATION_ALOHA.md

📦 Download dataset and Change dataset path

1. Download the dataset from ALOHA_Data

2. Modify constants.py Line 5 to your own dataset path

🚀 Model training and evaluation

Training script

cd ALOHA
bash script/train_eval.sh sim_insertion_human 20000 0 0
# bash script/train_eval.sh <task_name> <num_steps> <seed> <cuda_id>

Evaluation script

bash script/eval.sh sim_insertion_human 20000 0 0 0 
# bash script/train_eval.sh <task_name> <num_steps> <seed> <cuda_id> <ckpt_type>

RoboTwin

🔧 Installation

conda create -n RoboTwin python=3.10

1. Install the required packages for RoboTwin, see INSTALLATION_RoboTwin.md
2. Install the required packages for Cosmos-Tokenizer and download the checkpoints from Hugging Face, see Cosmos-Tokenizer
3. Install the required packages for policy deployment

pip install diffusers wandb ipdb gpustat dm_control omegaconf hydra-core==1.2.0 einops==0.4.1 diffusers==0.11.1 numba==0.56.4 moviepy imageio av matplotlib termcolor

📦 Data collection and preprocessing

cd RoboTwin
bash run_task.sh block_hammer_beat 0
# bash run_task.sh ${task_name} ${gpu_id}
python script/pkl2zarr_mypolicy.py block_hammer_beat D435 100
# python script/pkl2zarr_mypolicy.py ${task_name} ${head_camera_type} ${expert_data_num}

🚀 Model training and evaluation

Training script

cd policy/ACT-DP-TP
bash scripts/act_dp_tp/train.sh block_hammer_beat 0 0 
# bash scripts/train.sh ${task_name} ${gpu_id} ${seed}

Evaluation script

bash scripts/act_dp_tp/eval.sh block_hammer_beat 0 0 0
# bash scripts/eval.sh ${task_name} ${gpu_id} ${seed} ${ckpt_type}

🙏 Acknowledgements

Our project builds upon the following excellent repositories:

• ACT
• Cosmos-Tokenizer
• RoboTwin

We sincerely thank the authors for their inspiring work and open-source contributions.

Citation

If you find our work helpful, please cite us:

@misc{xu2025diffusionbasedimaginativecoordinationbimanual,
      title={Diffusion-Based Imaginative Coordination for Bimanual Manipulation}, 
      author={Huilin Xu and Jian Ding and Jiakun Xu and Ruixiang Wang and Jun Chen and Jinjie Mai and Yanwei Fu and Bernard Ghanem and Feng Xu and Mohamed Elhoseiny},
      year={2025},
      eprint={2507.11296},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2507.11296}, 
}

License

All the code, model weights, and data are licensed under MIT license.