Project
robomimic
Imitation-learning research library from Stanford/TRI with algorithms, datasets, and eval pipelines.
ProjectFeaturedBeginner
LeRobot
Hugging Face open-source stack for embodied AI with unified data formats and training — great for imitation learning and VLA experiments.
Why: One of the friendliest unified training and data-format frameworks today
Overview
LeRobot is Hugging Face’s open stack for robot learning. It standardizes dataset formats (LeRobotDataset), provides ready-made training scripts for imitation learning and VLA-style policies, and ships with evaluation utilities that lower the friction from “paper idea” to a runnable experiment.
Unlike research codebases that only reproduce a single paper, LeRobot aims to be a shared substrate: convert teleop logs, train ACT / Diffusion Policy / VLA variants, push datasets and models to the Hub, and share reproducible configs. That makes it especially valuable for newcomers and small labs that need a coherent default path rather than assembling five repos by hand.
In practice, teams use LeRobot as the data + training layer while pairing it with simulators (MuJoCo, ManiSkill, Isaac Lab) or low-cost real platforms (ALOHA-style arms). The Hub integration is a first-class feature: you can browse community datasets and checkpoints instead of rediscovering download scripts.
Who it is for
Beginners entering imitation learning / VLA; labs that want a shared data format; teams that already use Hugging Face Hub.
Key highlights
- Unified LeRobotDataset format with Hub upload/download
- Official recipes for popular imitation policies (ACT, Diffusion Policy, etc.)
- Strong docs and community examples for low-cost real robots
- Active development under Hugging Face robotics org
- Good default path from teleop → train → eval without custom plumbing
When to use
- You want one codebase to own data format + training loops
- You plan to share datasets/models on Hugging Face Hub
- You are starting with ACT / DP / simple VLA finetunes
- You prefer Python-first tooling over C++ control stacks
When not to use
- You need industrial real-time control or certified safety stacks
- Your primary work is pure sim RL at massive GPU scale (prefer Isaac Lab / ManiSkill envs)
- You only need a physics engine, not a learning framework
Getting started
- 1Clone the official repo and install the recommended Python env (follow README pin versions).
- 2Run a tiny example policy train on a Hub dataset to verify GPU / CUDA setup.
- 3Convert one of your teleop sessions into LeRobotDataset and visualize a few episodes.
- 4Train ACT or Diffusion Policy with an official config; log success rate, not only loss.
- 5Export a checkpoint and document camera FPS, action dim, and control frequency.
Install / setup
pip (see docs for extras)
pip install lerobot
clone
git clone https://github.com/huggingface/lerobot.git
Papers & reading
How it compares
Friendlier unified data+train path than assembling ACT/DP repos by hand; less specialized than robomimic for research ablations.
★ 1.5k
ProjectFeatured
openpi
Open training and inference code around π0 from Physical Intelligence for generalist robot policies.
Caveats & pitfalls
- API and dataset schema evolve quickly — pin commit hashes for papers.
- Real-robot success depends more on calibration and latency than on the training script.
- Stars and Hub popularity do not guarantee fitness for your embodiment.
Content reviewed 2026-07-23