Project
robosuite
Modular MuJoCo-based robot simulation framework for composing manipulation tasks and algorithms.
ProjectIntermediate
robomimic
Imitation-learning research library from Stanford/TRI with algorithms, datasets, and eval pipelines.
Overview
robomimic is a Stanford/TRI research library that packages imitation learning algorithms, datasets, and evaluation pipelines in one place. It became a reference implementation for comparing BC, BC-RNN, HBC, IRIS, and related methods on standardized robosuite tasks.
If you are writing an imitation learning paper and need reproducible baselines, robomimic remains a high-signal codebase. It is more research-oriented than product frameworks like LeRobot: expect configs, paper-aligned experiments, and less Hub-centric workflow.
Who it is for
Imitation learning researchers needing classic baselines; students reading the robomimic paper series.
Key highlights
- Canonical IL algorithm implementations
- Standardized datasets and eval protocols
- Tight integration with robosuite
- Strong paper reproducibility heritage
When to use
- You need paper-faithful IL baselines
- Your env stack is robosuite / MuJoCo
- You compare sequence models for demonstration learning
When not to use
- You want Hub-first modern VLA finetuning (LeRobot / OpenVLA)
- Your focus is real low-cost bimanual hardware day one
Getting started
- 1Install robomimic + robosuite as documented.
- 2Download a released dataset and run a BC baseline.
- 3Reproduce a table from the paper before modifying algorithms.
- 4Swap in your demos only after metrics match expectations.
Install / setup
git clone https://github.com/ARISE-Initiative/robomimic.git
Papers & reading
Caveats & pitfalls
- Less “productized” DX than newer frameworks.
- Some dependencies lag modern PyTorch defaults — pin carefully.
Content reviewed 2026-07-23