CtrlK
LeRobot is an open-source library for end-to-end robot learning developed by Hugging Face and a global community of contributors. Launched in May 2024 and led by Remi Cadene, a former staff scientist at Tesla Autopilot who helped build the first neural networks of the Optimus humanoid program, LeRobot was conceived as a kind of "Transformers for robotics": a PyTorch native toolkit bundling pretrained policies, standardized datasets, reference hardware designs, and simulation environments behind a single interface, then plugging the whole stack into the Hugging Face Hub for sharing [1] [2]. The project's mission is to lower the barrier to entry for robotics so that anyone with a laptop and a few hundred dollars of hardware can train, fine-tune, and deploy real-world manipulation policies, then upload their datasets and weights for others to build on. By April 2026, the GitHub repository had grown past 23,000 stars, the Hub hosted hundreds of community-contributed robotics datasets tagged lerobot, and the library had become the de facto entry point for hobbyist, academic, and increasingly industrial robot learning workflows [3] [4].
LeRobot's appeal comes from a tight loop between three things that historically lived in separate worlds: a unified LeRobotDataset format that makes a demonstration on a $150 desktop arm look the same to a model as a humanoid trajectory in a research lab; a growing zoo of policy implementations covering imitation learning, reinforcement learning, and modern vision-language-action models such as ACT, Diffusion Policy, TDMPC, VQ-BeT, Pi0, Pi0.5, SmolVLA, and NVIDIA's GR00T N1.5; and a set of reference robots, most famously the SO-100 and SO-101 low-cost 3D-printable arms, that let new users go from cloning the repository to training their first policy in an afternoon.
LeRobot's origin is tied to one hire. In March 2024, Hugging Face announced that it had recruited Remi Cadene from Tesla, where he had spent three years as a staff scientist on the Autopilot neural networks and the early perception models for the Optimus humanoid robot [5]. Hugging Face's pitch was to give him a small team and the freedom to do for robot learning what the Transformers library had done for natural language processing: package together the standard models, standardize the data format, and put the whole thing on the Hub.
The library's first public release went out on May 6, 2024, with an initial set of pretrained policies (ACT, Diffusion Policy, TDMPC), simulation environments adapted from existing research codebases, and a handful of community datasets [1] [6]. The response was immediate. The repository accumulated thousands of stars in the first weeks, and a Discord server filled with hobbyists asking how to wire up their first arms. Cadene used Twitter and YouTube to demonstrate the library on cheap hardware, including a series of posts in which he taught a sub-$300 arm to perform pick-and-place tasks after a few hundred demonstrations.
The single most consequential release of LeRobot's first year was not a model but a piece of hardware. In late 2024, Hugging Face partnered with The Robot Studio to publish the SO-100, a 6 degree-of-freedom 3D-printable arm with a bipartite gripper and a bill of materials around $150 per arm [7]. The design called for a leader and a follower so that users could collect demonstrations by puppeting the follower with the leader, capturing joint trajectories for training. Build instructions, STL files, and wiring guides were published openly, and distributors including Seeed Studio, Waveshare, Hiwonder, and ROBOTIS began selling kits within weeks. An upgraded SO-101 followed in April 2025, with simpler wiring, easier assembly, improved motors that reduced friction enough for the leader to hold its own weight, and a starting kit price around $100 [8] [9]. The SO arm series became the most cloned open robot in the world; by mid-2025 the bulk of community-contributed LeRobotDatasets had been recorded on either the SO-100 or the closely related Koch arm [10].
On April 14, 2025, Hugging Face announced its acquisition of Pollen Robotics, a French open-source robotics company building the Reachy humanoid platform since 2016 [11]. Hugging Face's chief scientist Thomas Wolf framed the deal publicly as a bet that "robotics could be the next frontier unlocked by AI" and that this frontier "should be open, affordable, and private," with Pollen co-founder Matthieu Lapeyre calling Hugging Face "a natural home for us to grow" [11] [12]. The deal made Reachy 2 the first commercial open-source humanoid powered end to end by LeRobot. Reachy 2, priced at $70,000 and already in use at Cornell and Carnegie Mellon, brought 7 degree-of-freedom human-inspired arms with the proprietary Orbita joint, a mobile base with omniwheels and LiDAR, and VR teleoperation as standard. The acquisition was Hugging Face's fifth and the first that involved physical product manufacturing rather than pure software.
On May 29, 2025, Hugging Face announced two more reference robots: HopeJR, a full-sized humanoid built with The Robot Studio with 66 actuated degrees of freedom and a target price around $3,000, and Reachy Mini, a tabletop expressive robot priced under $300 [13] [14]. HopeJR's blueprints and bill of materials were released openly on GitHub, with LeRobot integration handled by Martino Russi on the LeRobot team. The exoskeleton-style leader rig used to teleoperate HopeJR provided 1:1 joint mapping and required no soldering, an extension of the philosophy that had made the SO arms so accessible. Combined with existing support for Mobile ALOHA, the Hello Robot Stretch, and Unitree's G1, the LeRobot stack now spanned almost every major form factor relevant to embodied AI research.
LeRobot is written almost entirely in Python on top of PyTorch and is published under the Apache 2.0 license, installable via pip as the lerobot package. The repository is organized around a small set of orthogonal abstractions: datasets, policies, robots, environments, and configuration. Each lives in its own subpackage with a consistent interface, so swapping one component (for example replacing a Diffusion Policy with a Pi0 policy) does not require rewriting the others. The Hub is treated as a first-class storage backend; the same code can pull pretrained weights from lerobot/pi0_base or push a fine-tuned checkpoint back. Significant engineering went into the data path because robotics workloads mix high-frequency tabular data (joint positions, gripper widths, end-effector poses at 30 Hz to 1000 Hz) with large video streams from cameras at 30 to 60 fps. The LeRobotDataset format is the team's answer; versions 1, 2, 2.1, and 3.0 all shipped within eighteen months. Version 3.0 (late 2025) switched from one-file-per-episode to a chunked multi-episode-per-file layout that scales to millions of episodes and supports streaming directly from the Hub [15].
LeRobot's policy zoo is one of the project's main attractions. Rather than reimplement each algorithm from scratch, the team standardized the interface so a research codebase can be ported in a few hundred lines and immediately gain access to the dataset format, simulation harness, and Hub integration. As of 2026, the library ships reference implementations spanning imitation learning, reinforcement learning, and large vision-language-action models.
| Policy | Family | Origin | Notes |
|---|---|---|---|
| ACT | Imitation learning | Tony Zhao et al. (Stanford, 2023) | Action Chunking with Transformers; default first-try policy in LeRobot tutorials, lightweight and fast to train [16] |
| Diffusion Policy | Imitation learning | Cheng Chi et al. (Columbia, 2023) | Generates action chunks via denoising; strong on multimodal demonstration distributions [17] |
| VQ-BeT | Imitation learning | Behavior Transformer family (NYU, 2024) | Tokenizes continuous actions with a Residual VQ-VAE, trains a GPT-style transformer on the tokens |
| Multitask DiT | Imitation learning | LeRobot team | Diffusion Transformer trained jointly across many task conditions |
| TDMPC | Reinforcement learning | Nicklas Hansen et al. (UCSD, 2022) | Model-based RL with temporal-difference predictions, often used in simulation benchmarks |
| HIL-SERL | Reinforcement learning | UC Berkeley | Human-in-the-loop sample-efficient RL for short manipulation tasks |
| Pi0 | Vision-language-action | Physical Intelligence (2024) | Flow-matching VLA producing 50 Hz action trajectories; ported from Physical Intelligence's open OpenPI release [18] |
| Pi0-FAST | Vision-language-action | Physical Intelligence (2025) | Autoregressive variant of Pi0 with frequency-space action tokenization |
| Pi0.5 | Vision-language-action | Physical Intelligence (2025) | Open-world generalization VLA, co-trained on heterogeneous robot, web, and high-level semantic data [19] |
| SmolVLA | Vision-language-action | LeRobot team (2025) | 450M parameter VLA pretrained on 10M frames from 487 community datasets; runs on a single consumer GPU [20] |
| GR00T N1.5 | Vision-language-action | NVIDIA (2025) | Open foundation model for humanoids; integrated via LeRobot dataset format and inference path [21] |
| XVLA | Vision-language-action | Multiple contributors | Cross-embodiment VLA with shared action heads |
ACT remains the most-used policy in tutorials, small enough to train in a few hours on a single consumer GPU and robust enough to learn pick-and-place from a few hundred demonstrations on an SO-100. Diffusion Policy is preferred when the demonstration set contains multimodal behaviors because the diffusion sampling head represents multimodality directly. The VLAs are heavier and require either large multi-robot datasets or strong pretraining, but they are the route through which LeRobot users access generalist behavior across tasks and embodiments. SmolVLA, released by the LeRobot team in mid-2025, was particularly important because it was the first VLA designed to be cheap enough that a hobbyist could fine-tune it on a single consumer GPU. Pretrained on 10 million frames curated from 487 community datasets, it outperformed much larger VLAs and the ACT baseline on simulation benchmarks (LIBERO, Meta-World) and real SO-100/SO-101 tasks [20]; its asynchronous inference path gives roughly 30 percent faster response and twice the task throughput compared to synchronous inference.
LeRobot's hardware support roster has grown from two arms at launch to more than a dozen platforms by 2026, spanning low-cost desktop arms, bimanual rigs, mobile manipulators, and full humanoids. Each supported robot has a driver in the lerobot package that exposes a standard set of methods for connecting, sending commands, reading state, and recording teleoperated trajectories into a LeRobotDataset. Many can run in either real-world or simulation mode with only a configuration change.
| Robot | Form factor | Approximate cost | Origin | Notes |
|---|---|---|---|---|
| SO-100 | 6-DoF desktop arm | $150 BOM | Hugging Face and TheRobotStudio | Bipartite gripper, leader and follower configuration, 3D-printable [7] |
| SO-101 | 6-DoF desktop arm | $100 to $130 | Hugging Face and TheRobotStudio | Improved wiring and motors over SO-100, easier assembly [8] [9] |
| Koch v1.1 | 6-DoF desktop arm | $250 to $400 | Alexander Koch and Jess Moss | Earlier reference design, still widely used for teleoperation |
| Moss v1 | 6-DoF desktop arm | varies | Community | Slated for deprecation as SO-101 takes over |
| LeKiwi | Mobile manipulator | $500 to $800 | SIGRobotics-UIUC | Three-wheeled base with embedded Raspberry Pi, often paired with SO arm |
| ALOHA | Bimanual stationary | ~$20,000 | Stanford (Tony Zhao et al.) | 14-DoF bimanual teleoperation kit, classic VLA benchmark hardware |
| ALOHA 2 | Bimanual stationary | varies | Stanford and Google DeepMind | Improved ergonomics and reliability over original ALOHA |
| Mobile ALOHA | Bimanual mobile | ~$32,000 | Stanford | Whole-body teleoperation with a mobile base |
| Stretch | Mobile manipulator | ~$25,000 | Hello Robot | Telescoping arm on a mobile base, supports dex teleop |
| Trossen WidowX AI | 7-DoF arm | varies | Trossen Robotics | Industrial-grade arm with native LeRobot integration |
| Reachy 2 | Humanoid (waist-up) | $70,000 | Pollen Robotics | First commercial open-source humanoid powered end-to-end by LeRobot [11] |
| Reachy Mini | Tabletop expressive | under $300 | Hugging Face and Pollen | Small developer-focused unit announced May 2025 [13] |
| HopeJR | Full humanoid | ~$3,000 | Hugging Face and TheRobotStudio | 66-DoF humanoid with exoskeleton leader for teleoperation [13] |
| OpenArm 101 | 6-DoF research arm | varies | Open Arm community | ALOHA-compatible data format, integrates with LeRobot |
| HopeJR Hand | Dexterous gripper | ~$500 | TheRobotStudio | Anthropomorphic hand for HopeJR or as standalone end effector |
| OMX | Open Manipulator X | ~$1,500 | ROBOTIS | Educational arm with native ROBOTIS LeRobot drivers |
| Earth Rover | Outdoor mobile | varies | Earth Rover Challenge | Outdoor wheeled platform for navigation policies |
| Unitree G1 | Full humanoid | ~$16,000 | Unitree Robotics | Bipedal humanoid with LeRobot inference support |
The combination of cheap arms, mobile manipulators, and full humanoids in a single library is unusual; most academic robot learning codebases are written for a single platform and require substantial rewriting to port. LeRobot's standardization on the LeRobotDataset format and a uniform policy interface means that a model trained on one platform can in principle be evaluated on another without changing the policy code, although cross-embodiment transfer remains an open research problem.
The LeRobotDataset is the standardized format that ties the rest of the library together. It addresses the dual nature of robot learning data: a steady stream of low-dimensional, high-frequency tabular data (joint positions, velocities, end-effector poses, gripper widths, action commands) interleaved with high-bandwidth video streams from one or more cameras. Naive solutions such as one-file-per-episode fall apart on real workloads where datasets contain hundreds of thousands of episodes and many terabytes of video.
The v3.0 layout, released in late 2025, organizes data into three components [15]. Tabular data is stored in chunked Apache Parquet files with multiple episodes per file, allowing fast memory-mapped or streaming access. Visual data is stored as MP4 video files, with frames from many episodes concatenated under a directory tree organized by camera key and chunk index. Metadata lives in JSON or Parquet files: meta/info.json holds schema, frame rates, and shapes, meta/stats.json holds normalization statistics, meta/tasks.jsonl maps task descriptions to indices, and meta/episodes/ contains per-episode boundaries resolved through metadata rather than filenames, which is what allows the format to scale.
A dataset can be loaded in a single line through LeRobotDataset("yaak-ai/L2D-v3"), with a delta_timestamps argument that lets users request stacked observations across time windows for policies like Diffusion Policy and ACT. A StreamingLeRobotDataset variant pulls data on demand from the Hub without ever downloading the full dataset locally, necessary for community datasets that exceed several hundred gigabytes [15]. The format is compatible with the standard PyTorch DataLoader. The Hub now hosts thousands of LeRobotDatasets tagged lerobot; most were recorded on SO-100 and Koch arms, but the catalog also includes the yaak-ai L2D-v3 driving dataset (10,000+ episodes, ~500 GB), bimanual ALOHA datasets, Mobile ALOHA whole-body datasets, and the GR00T N1 humanoid dataset NVIDIA released as part of its larger open physical AI dataset push [21] [15].
LeRobot wraps several existing simulation environments in a common interface so that policies can be trained, evaluated, and benchmarked without physical hardware. The PushT 2D pushing environment, originally introduced as a small Diffusion Policy benchmark, ships as the simplest test bed for new contributors. ALOHASim provides the bimanual MuJoCo simulation that mirrors the real ALOHA hardware, ManiSkill provides a broader catalog of dexterous manipulation tasks built on the SAPIEN physics engine, and PandaGym wraps the Franka Emika Panda arm in an OpenAI Gym style interface. RoboCasa, contributed by the Stanford BEHAVIOR team, adds large-scale procedurally generated household kitchens for long-horizon manipulation. LIBERO and Meta-World, both standard imitation learning benchmarks, can be loaded directly, and SmolVLA's evaluation table reports results on both [20]. For users who want to bridge simulation and the real world, LeRobot integrates with NVIDIA Isaac Sim and Isaac Lab through the GR00T pipeline: the GR00T blueprint for synthetic manipulation motion generation produces synthetic training data in the LeRobotDataset format, which is then used to post-train the GR00T N1 model without ever touching physical hardware [21].
LeRobot's growth has been driven as much by community as by Hugging Face's own engineering. The Discord passed 10,000 members by the end of 2025, the YouTube ecosystem hosts hundreds of independent build videos, and the lerobotdepot community repository tracks open-source hardware compatible with the library. University courses at Stanford, MIT, CMU, EPFL, and many other institutions have integrated SO-100 builds into curricula, and industry partnerships with Trossen Robotics, Hello Robot, ROBOTIS, Seeed Studio, Waveshare, and Hiwonder have produced commercial kit versions of the open hardware designs. By 2026, hundreds of community-tagged LeRobotDatasets and dozens of community-fine-tuned policies were available for direct download. SmolVLA's pretraining on 487 community datasets [20] and NVIDIA's GR00T N1 dataset release [21] cemented the LeRobotDataset format as the lingua franca for sharing robot data publicly.
The table below summarizes the people and partner organizations most associated with the project.
| Contributor | Role | Affiliation |
|---|---|---|
| Remi Cadene | Principal research scientist; project lead | Hugging Face (formerly Tesla Autopilot) |
| Thomas Wolf | Co-founder and chief scientist; LeRobot sponsor | Hugging Face |
| Clement Delangue | CEO; public face of the LeRobot strategy | Hugging Face |
| Matthieu Lapeyre | Reachy lead; brought Pollen Robotics into LeRobot | Pollen Robotics, now Hugging Face |
| Pierre Rouanet | Co-founder of Pollen Robotics | Pollen Robotics, now Hugging Face |
| Martino Russi | Hardware lead for HopeJR and SO arm series | Hugging Face |
| Alexander Koch | Designer of Koch arm reference | Independent |
| Jess Moss | Maintainer of Koch v1.1 | Independent |
| The Robot Studio | SO-100, SO-101, HopeJR co-design | Independent design firm |
| Tony Zhao | Co-author of ACT and ALOHA | Stanford / Physical Intelligence |
| Sergey Levine | Co-PI on ALOHA, Pi0, Pi0.5 | Physical Intelligence and UC Berkeley |
| Karol Hausman | Pi0 lead | Physical Intelligence |
| Chelsea Finn | Pi0 advisor | Stanford / Physical Intelligence |
| Cheng Chi | Diffusion Policy lead | Columbia, now Stanford |
| NVIDIA Isaac Lab team | GR00T integration and Isaac Sim bridge | NVIDIA |
The partnership with Physical Intelligence, the well-funded robotics startup founded in 2024 by Karol Hausman, Sergey Levine, Chelsea Finn, and others, has been one of the most consequential. Physical Intelligence open-sourced its Pi0 VLA under the name OpenPI in late 2024, and the LeRobot team adapted it within weeks. Pi0 introduced flow matching as the action head for a 3 billion parameter VLA, producing smooth 50 Hz trajectories and demonstrating strong cross-task and cross-robot generalization [18]. The follow-up Pi0.5, released in April 2025, focused on open-world generalization through co-training on heterogeneous robot, web, and high-level semantic data, and is used in LeRobot as the standard high-end baseline for tasks that require zero-shot transfer to new objects and environments [19].
NVIDIA's GR00T project, announced as a foundation model for humanoids at GTC 2024 and expanded at GTC 2025 with the launch of GR00T N1, treats LeRobot as the open distribution channel for both data and weights. GR00T N1 and its successors N1.5 and N1.7 are distributed on the Hugging Face Hub under the nvidia/GR00T-N1-2B family, the training data is published in the LeRobotDataset format, and Isaac Lab natively supports LeRobot for converting synthetic motion-generation output into trainable datasets [21] [22]. Jensen Huang's GTC 2025 keynote demonstrated 1X's humanoid performing autonomous domestic tidying using a post-trained policy built on GR00T N1, and the Fourier GR-1 humanoid was shown executing a multi-step pick-transfer-pass sequence using the same model.
TheRobotStudio, a French independent design firm responsible for several open-source robot designs since the early 2010s, has been Hugging Face's primary hardware co-design partner. The SO-100, SO-101, HopeJR, and HopeJR Hand all originated from joint design work between The Robot Studio's engineers and the LeRobot software team. The collaboration publishes designs under permissive open licenses and does not enforce IP on the bills of materials, which is what has allowed third-party manufacturers to ship pre-printed kits at low margins.
LeRobot has been received as a turning point for open-source robotics in roughly the way that Hugging Face's Transformers library was for natural language processing. Press coverage at launch in May 2024 emphasized the unusual move of a major AI company entering a domain dominated by closed industrial vendors and academic codebases [1] [6]. By the end of 2025, technology publications were citing the SO-100 as an example of how open-source hardware combined with shared data could undercut closed industrial pricing, and the broader robot learning community had converged on LeRobotDataset as the de facto data format. Papers that release demonstration data now routinely include a LeRobotDataset version on the Hub, and workshops on imitation learning and generalist robotic policies at NeurIPS, CoRL, and ICRA increasingly cite the library as the standard baseline [23]. Industry reception has been more divided: several established manufacturers have integrated LeRobot drivers (Trossen, Hello Robot, ROBOTIS), while others have positioned themselves as offering closed alternatives. The Pollen acquisition and the launch of HopeJR signaled that Hugging Face is willing to compete directly in the hardware market when no existing commercial offering meets the affordability and openness criteria the team values.
LeRobot is still a young project. Cross-embodiment transfer, in which a policy trained on one robot generalizes to another with a different action space and morphology, remains an unsolved research problem; the VLAs in the policy zoo make progress on this front but are far from a full solution. Sim-to-real transfer is still expensive and often requires domain-randomization engineering that the current tutorials do not cover well. The bundled simulators are research benchmarks rather than industrial-grade systems, and serious sim-to-real workflows still tend to involve external tools such as Isaac Sim or MuJoCo MJX configured outside the LeRobot config system. The policy zoo has also grown faster than the documentation, and a more guided onboarding flow has been a frequent request from the Discord community.