LeRobot is an open-source project designed to democratize AI for robotics by lowering the barrier to entry through standardized models, datasets, and tools in PyTorch, targeting real-world applications. Its core methodology revolves around a three-pronged approach: a hardware-agnostic control interface, a standardized data format, and implementations of state-of-the-art machine learning policies.

Firstly, LeRobot introduces a unified Robot & Control interface, conceptualized as a Robot Python class. This class effectively decouples control logic from hardware specifics, allowing for a single programmatic interface to interact with diverse robotic platforms. The interface standardizes fundamental robotic operations such as connect(), get_observation(), and send_action(). For example, $obs = robot.get_observation()$ retrieves sensor data, and robot.send_action(action) dispatches control commands. This abstraction supports a wide range of hardware, including low-cost arms (e.g., SO100, OpenARM), humanoids (Unitree G1), and various teleoperation devices (Gamepads, Keyboards, Phones), while remaining extensible for custom robot implementations by adhering to the Robot interface.

Secondly, to address the prevalent data fragmentation problem in robotics, LeRobot proposes the LeRobotDataset format. This standardized, scalable format combines synchronized MP4 videos (or image sequences) for visual observations with Parquet files for state and action data. This structure facilitates efficient storage, streaming, and visualization of massive robotic datasets, which are hosted and accessible directly from the Hugging Face Hub. The LeRobotDataset class provides functionalities for seamless data access, such as $dataset = LeRobotDataset("lerobot/aloha_mobile_cabinet")$, allowing direct indexing (e.g., dataset[episode_index]['action']) for retrieving specific modalities. Furthermore, the framework includes robust tools for dataset manipulation, including deleting episodes, splitting datasets by indices or fractions, adding or removing features, and merging multiple datasets.

Thirdly, LeRobot provides implementations of State-of-the-Art (SoTA) Models, or policies, in pure PyTorch, covering various machine learning paradigms applicable to robotics. These policies are designed for direct training and deployment, demonstrating real-world transfer capabilities. The supported categories include:

• Imitation Learning: Policies like ACT (Action Chunking with Transformers), Diffusion Models, and VQ-BeT.
• Reinforcement Learning: Including HIL-SERL (Hierarchical Imitation Learning with Structured Exploration and Regularization Learning), with upcoming support for TDMPC (Temporal Difference Model Predictive Control) and QC-FQL (Q-value Constrained Fitted Q-Learning).
• Vision-Language-Action (VLA) Models: Such as Pi0.5, GR00T N1.5, SmolVLA, and XVLA.

Training these policies is streamlined via command-line scripts (e.g., $lerobot-train --policy=act --dataset.repo_id=lerobot/aloha_mobile_cabinet$). Similar to the hardware interface, the policy framework is extensible, allowing researchers and practitioners to implement their own custom policies and share them on the Hugging Face Hub. For evaluation, LeRobot offers a unified script (lerobot-eval) compatible with standard benchmarks like LIBERO and MetaWorld, enabling rigorous assessment of policy performance in both simulation and real-world environments.

LeRobot fosters an open-source ecosystem, leveraging the Hugging Face Hub for dataset and model distribution, and promoting community contributions for feature development, documentation, and bug fixes.