A software framework integrating various imitation learning methods and benchmark environments for robotic manipulation.
Provides easy-to-use baselines for policy training, evaluation, and deployment.
RoboManipBaselines_VideoVer200.mp4
RoboManipBaselines_VideoVer100.mp4
Quick Start
Start collecting data in the MuJoCo simulation, train your model, and rollout the ACT policy in just a few steps!
See the Quick Start Guide.
Installation
Follow our step-by-step Installation Guide to get set up smoothly.
Policies
We provide several powerful policy architectures for manipulation tasks:
- MLP: Simple feedforward policy
- SARNN: Recurrent policy for sequential data
- ACT: Transformer-based action chunking policy
- MT-ACT: Multi-task Transformer-based imitation policy
- Diffusion Policy: Diffusion-based imitation policy
- 3D Diffusion Policy: Diffusion-based policy with 3D point cloud input
- Flow Policy: Flow-matching-based policy with 3D point cloud input
- ManiFlow Policy: Flow-matching and consistency-based policy with 2D/3D vision
Data
- Dataset List: Pre-collected expert demonstration datasets
- Learned Parameters: Trained model checkpoints and configs
- Data Format: Description of the custom RMB data format used in RoboManipBaselines
- Point Cloud Preprocessing: Data preprocessing for 3D point cloud policies
Teleoperation
Use your own teleop interface to collect expert data.
See Teleop Tools for more info.
- Multiple SpaceMouse: Setup multiple SpaceMouse for high-degree-of-freedom robots
Environments
Explore diverse manipulation environments:
- Environment Catalog: Overview of all task environments
- Env Setup: Installation guides per environment
- How to Add a New Environment: Guide for adding a custom environment
- MuJoCo Tactile Sensor: Guide for using tactile sensors in MuJoCo environments
Miscellaneous
Check out Misc Scripts for standalone tools and utilities.
Evaluation Results
See Benchmarked Performance across environments and policies.
Contributing
We welcome contributions!
Check out the Contribution Guide to get started.
License
This repository is licensed under the BSD 2-Clause License, unless otherwise stated.
Please check individual files or directories (especially third_party and assets) for specific license terms.
Citation
If you use RoboManipBaselines in your work, please cite our paper:
title={RoboManipBaselines: A Unified Framework for Imitation Learning in Robotic Manipulation across Real and Simulated Environments},
author={Murooka, Masaki and Motoda, Tomohiro and Nakajo, Ryoichi and Oh, Hanbit and Makihara, Koshi and Shirai, Keisuke and Domae, Yukiyasu},
journal={arXiv preprint arXiv:2509.17057},
year={2025}
}