LeRobot: The Open-Source Robot Learning Library Explained

What Is LeRobot?

LeRobot is a Python library maintained by Hugging Face that provides end-to-end infrastructure for robot learning research. It handles four distinct concerns: recording robot demonstrations from real hardware, storing and versioning those demonstrations in a standardized dataset format, training state-of-the-art imitation learning and reinforcement learning policies on those datasets, and deploying trained policies back to real hardware for evaluation. Each of these capabilities is modular — you can use LeRobot for data storage only, or for full-stack training and deployment.

The project lives on GitHub under the huggingface organization and as of early 2026 has accumulated tens of thousands of stars and hundreds of contributors. Datasets can be published to and downloaded from the Hugging Face Hub, giving LeRobot a growing shared dataset repository that functions as a community data marketplace.

Supported Algorithms

LeRobot ships with native implementations of three policy classes. ACT (Action Chunking with Transformers) is the primary algorithm for fine-grained manipulation tasks: it uses a transformer-based CVAE architecture with temporal ensembling and is the recommended starting point for most dexterous manipulation workflows. Diffusion Policy implements both the CNN-based and transformer-based variants of denoising diffusion for action prediction, which excels on tasks with multimodal action distributions. TDMPC2 (Temporal Difference Model Predictive Control) is a model-based RL algorithm that learns both a world model and a policy, offering sample-efficient training when a simulation environment is available.

Each algorithm is implemented in PyTorch with standard training scripts, hydra configuration management, and Weights & Biases integration for experiment tracking. Switching between algorithms is a single config file change, making comparative evaluation straightforward.

Supported Hardware

Out-of-the-box hardware integrations include ALOHA bimanual systems (ViperX arms), Koch arms, SO-100 and SO-101 low-cost arms, Lekiwi mobile platforms, and several Dynamixel servo-based custom arms. The OpenArm, available through SVRC's store, has native LeRobot support for teleoperation recording and policy deployment. Adding a new hardware platform requires implementing a Robot interface class with methods for reading joint state and sending joint commands — typically a few hundred lines of code for a new arm.

Camera support covers USB cameras via OpenCV, Intel RealSense depth cameras, and webcam arrays. The recording system handles multi-camera synchronization with software timestamping and supports configurable frame rates and resolutions per camera.

The LeRobot Dataset Format

LeRobot stores datasets as HDF5 files with a standardized episode structure. Each episode contains arrays for observations (images compressed as video streams, joint states as float32 arrays), actions (joint position targets), timestamps, and annotations (task language strings, success flags). Metadata files describe the robot configuration, camera calibration, and dataset statistics. This format is designed to be self-describing and portable: a dataset recorded on one machine can be loaded for training on any other machine without modification.

Datasets published to the Hugging Face Hub include a dataset card with statistics, task descriptions, and usage examples. This makes it easy to discover and reuse datasets from the community, reducing the data collection burden for common tasks.

Getting Started with LeRobot

Installation requires Python 3.10+ and PyTorch. The recommended first project is recording 50 demonstrations of a simple pick-and-place task on any supported hardware, then training an ACT policy on those demonstrations. LeRobot's documentation provides a full walkthrough from hardware setup through policy evaluation. For teams without their own hardware, SVRC offers robot leasing through our leasing program specifically to enable LeRobot-based data collection.

SVRC Export Compatibility

SVRC's data platform exports datasets in native LeRobot HDF5 format, with all required metadata fields populated from the recording session. Datasets collected through SVRC's facility or data collection services arrive ready for training — no conversion step required. SVRC also provides dataset hosting on the Hugging Face Hub for research partners who want to share their data publicly. If you have questions about integrating SVRC data with your LeRobot training pipeline, the SVRC engineering team is available to help.