README.md

metadata

task_categories:
  - robotics
tags:
  - reinforcement-learning
  - robomimic

From Prior to Pro: Efficient Skill Mastery via Distribution Contractive RL Finetuning (DICE-RL)

Project Website | Paper | GitHub

This repository contains the datasets used in DICE-RL, a framework that uses reinforcement learning as a "distribution contraction" operator to refine pretrained generative robot policies. The data includes both pretraining data (for Behavior Cloning) and finetuning data (for DICE-RL) across various Robomimic environments.

Dataset Structure

The datasets are provided in numpy format, and each folder typically contains train.npy and normalization.npz. The data is organized following this structure:

data_dir/
└── robomimic
    ├── {env_name}-low-dim
    │   ├── ph_pretrain
    │   └── ph_finetune
    └── {env_name}-img
        ├── ph_pretrain
        └── ph_finetune

• ph_pretrain: Contains the datasets used for pretraining the BC policies.
• ph_finetune: Contains the datasets used for finetuning the DICE-RL policies. These are similar to the pretraining sets but with trajectories truncated to ensure value learning consistency between offline and online data (truncated to have exactly one success at the end).
• low-dim: State-based observations.
• img: High-dimensional pixel (image) observations.

Usage

You can download the datasets using the scripts provided in the GitHub repository:

bash script/download_hf.sh

For more details on generating your own data or processing raw Robomimic datasets, please refer to the project's dataset processing guide.

Citation

@article{sun2026prior,
  title={From Prior to Pro: Efficient Skill Mastery via Distribution Contractive RL Finetuning},
  author={Sun, Zhanyi and Song, Shuran},
  journal={arXiv preprint arXiv:2603.10263},
  year={2026}
}