qinghangliu/DADP

This repository contains the datasets used in the paper DADP: Domain Adaptive Diffusion Policy.

Project Page | GitHub Repository

Dataset Description

DADP (Domain Adaptive Diffusion Policy) is a framework for learning domain-adaptive policies that can generalize to unseen transition dynamics. The datasets provided include trajectories for various locomotion and manipulation benchmarks:

• Locomotion: Ant, Walker2d, HalfCheetah, Hopper.
• Manipulation: Adroit (Door, Relocate).

The manipulation datasets are sourced from the ODRL (Off-Dynamics RL) project and are noted to be smaller in size and near-random in quality compared to the locomotion environments.

Usage

As per the official repository, these datasets are intended to be used with the Minari framework. Once you have downloaded the datasets, extract and move them into your local Minari datasets directory (typically ~/.minari/datasets/).

The expected directory structure is:

~/.minari/
└── datasets/
    ├── RandomAnt/
    ├── RandomWalker2d/
    ├── Adroit/
    └── ...

For detailed instructions on training and evaluation, please refer to the official GitHub README.

Citation

@article{liu2024dadp,
  title={DADP: Domain Adaptive Diffusion Policy},
  author={Liu, Qinghang and others},
  journal={arXiv preprint arXiv:2602.04037},
  year={2024}
}