Falcon

Official codebase for "Falcon: Fast Visuomotor Policies via Partial Denoising". This repository is based on Diffusion Policy.

Installation

See the original Diffusion Policy repo for installation.

Contributions

Our contributions to the repo are:

• The definition of state-based Falcon, in diffusion_policy/policy, in the file falcon_unet_lowdim_policy.py.
• Some utils about Falcon in diffusion_policy/common/falcon_util.py
• The modified model architecture in diffusion_policy/model/diffusion/parallel_conditional_unet1d.py
• The evaluation in eval_falcon.py

Evaluation

We only support evaluation with single env. For streaming diffusion policy baseline, please refer to their codebase.

Download Checkpoints

Download checkpoints from the official Diffusion Policy repo

Eval with ddpm

checkpoint_path=<checkpoint_path>
device="cuda:0"
seed=<seed>
output_dir="eval_dir/can_ph/diffusion_policy_cnn/ddpm/seed_${seed}"
epsilon=<threshold>
delta=<exploration_rate>
k_min=<minimum_noise_level>
max_buffersz=<buffer_size>
temperature=<temperature>
python eval.py --checkpoint $checkpoint_path \
                -o $output_dir \
                -d $device \
                --scheduler ddpm \
                --seed $seed

Eval with ddpm+falcon

checkpoint_path=<checkpoint_path>
device="cuda:0"
seed=<seed>
output_dir="eval_dir/can_ph/diffusion_policy_cnn/falconddpm/seed_${seed}"
epsilon=<threshold>
delta=<exploration_rate>
k_min=<minimum_noise_level>
max_buffersz=<buffer_size>
temperature=<temperature>
python eval.py --checkpoint $checkpoint_path \
                -o $output_dir \
                -d $device \
                --scheduler falconddpm \
                --epsilon $epsilon \
                --delta $delta \
                --k_min $k_min \
                --max_buffersz $max_buffersz \
                --temperature $temperature \
                --seed $seed

Citation

If you find our work useful, please consider citing our paper:

@article{chen2025fast,
  title={Fast Visuomotor Policies via Partial Denoising},
  author={Chen, Haojun and Liu, Minghao and Ma, Xiaojian and Ma, Zailin and Wu, Huimin and Ma, Chengdong and Chen, Yuanpei and Zhong, Yifan and Wang, Mingzhi and Li, Qing and others},
  journal={arXiv preprint arXiv:2503.00339},
  year={2025}
}

Acknowledgements

We thank the authors of Diffusion Policy and Streaming Diffusion Policy for sharing their codebase.