umi-on-air_checkpoints

UMI-on-Air - Pre-trained Checkpoints

Pre-trained diffusion policy checkpoints for UMI-on-Air: Embodiment-Aware Guidance for Embodiment-Agnostic Visuomotor Policies.

UMI-on-Air is a framework for embodiment-aware deployment of embodiment-agnostic manipulation policies. It leverages diverse, unconstrained human demonstrations to train generalizable visuomotor policies that can be adapted to constrained robotic embodiments at inference time.

📄 Paper: arXiv | Hugging Face Papers

🌐 Project Website: umi-on-air.github.io

💻 Code: GitHub

Checkpoints Included

Usage

1. Download and Extract

# Download and extract the checkpoints
wget https://huggingface.co/LeCAR-Lab/umi-on-air_checkpoints/resolve/main/checkpoints.tar.gz
tar -xzf checkpoints.tar.gz

Each checkpoint folder contains:

• checkpoints/latest.ckpt - Model weights
• normalizer.pkl - Data normalization parameters
• .hydra/config.yaml - Training configuration

2. Policy Evaluation

After setting up the environment according to the official repository, you can evaluate a policy in simulation with Embodiment-Aware Diffusion Policy (EADP) guidance.

For example, to evaluate the Unmanned Aerial Manipulator (UAM) on the cabinet task with guidance:

cd am_mujoco_ws/policy_learning
python imitate_episodes.py \
    --task_name uam_cabinet \
    --num_rollouts 30 \
    --guidance 1.5 \
    --use_3d_viewer

Citation

If you use these checkpoints, please cite our work:

@misc{gupta2025umionairembodimentawareguidanceembodimentagnostic,
      title={UMI-on-Air: Embodiment-Aware Guidance for Embodiment-Agnostic Visuomotor Policies}, 
      author={Harsh Gupta and Xiaofeng Guo and Huy Ha and Chuer Pan and Muqing Cao and Dongjae Lee and Sebastian Scherer and Shuran Song and Guanya Shi},
      year={2025},
      eprint={2510.02614},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2510.02614}, 
}