README.md

metadata

license: mit
task_categories:
  - robotics
tags:
  - reinforcement-learning
  - imitation-learning
  - robot-manipulation

Beyond Action Residuals: Real-World Robot Policy Steering via Bottleneck Latent Reinforcement Learning (ZPRL)

Project Page | Paper | GitHub

This repository contains the datasets used in the paper "Beyond Action Residuals: Real-World Robot Policy Steering via Bottleneck Latent Reinforcement Learning".

Dataset Description

These datasets are prepared for training and evaluating robot manipulation policies using the ZPRL framework. They are based on the Robomimic Multi-Human (MH) dataset and include the following tasks:

• Can
• Square
• Transport

Each dataset consists of 100 trajectories randomly sampled from the original MH dataset. The data has been processed to include:

1. Image Observations: Rendered image observations at the resolution required for training.
2. Absolute Actions: Delta actions from the original datasets have been converted into absolute actions to facilitate training with flow-matching policies.

File Structure

The datasets are provided in .hdf5 format. The directory structure is organized as follows:

robomimicv030
├── can
│   └── mh
│       └── image_v141_subset_abs.hdf5
├── square
│   └── mh
│       └── image_v141_subset_abs.hdf5
└── transport
    └── mh
        └── image_v141_subset_abs.hdf5

Citation

If you find this dataset or the associated code useful, please consider citing the following paper:

@misc{yu2026zprl,
      title={Beyond Action Residuals: Real-World Robot Policy Steering via Bottleneck Latent Reinforcement Learning},
      author={Dongjie Yu and Kun Lei and Zhennan Jiang and Jia Pan and Huazhe Xu},
      year={2026},
      eprint={2605.19919},
      archivePrefix={arXiv},
      url={https://arxiv.org/abs/2605.19919},
}