This model implements a Mixture of Diffusion Experts architecture for robotic manipulation, combining transformer-based backbone with noise-only expert routing. For faster inference, we can precache the chosen expert for each timestep to reduce computation time.

The model has been pretrained on a subset of OXE for 300k steps and finetuned for can be finetuned for downstream tasks.

Model Details

Architecture

Base Architecture: MoDE with custom Mixture of Experts Transformer
Vision Encoder: ResNet-50 with FiLM conditioning finetuned from ImageNet
EMA: Enabled
Action Window Size: 10
Sampling Steps: 5 (optimal for performance)
Sampler Type: DDIM

Input/Output Specifications

Inputs

RGB Static Camera: (B, T, 3, H, W) tensor
RGB Gripper Camera: (B, T, 3, H, W) tensor
Language Instructions: Text strings

Outputs

Action Space: (B, T, 7) tensor representing delta EEF actions

Usage

python

obs = {
    "rgb_obs": {
        "rgb_static": static_image,
        "rgb_gripper": gripper_image
    }
}
goal = {"lang_text": "pick up the blue cube"}
action = model.step(obs, goal)

Training Details

Configuration

Optimizer: AdamW
Learning Rate: 0.0001
Weight Decay: 0.05

Citation

If you found the code usefull, please cite our work:

bibtex

@misc{reuss2024efficient,
    title={Efficient Diffusion Transformer Policies with Mixture of Expert Denoisers for Multitask Learning},
    author={Moritz Reuss and Jyothish Pari and Pulkit Agrawal and Rudolf Lioutikov},
    year={2024},
    eprint={2412.12953},
    archivePrefix={arXiv},
    primaryClass={cs.LG}
}

License

This model is released under the MIT license.