LeRobot

LeRobot is an open-source, PyTorch-native library for end-to-end real-world robot learning developed by Hugging Face and a global community of contributors, often described as the "Transformers for robotics": a single toolkit that bundles pretrained policies, standardized datasets, low-cost reference hardware, and simulation environments, all wired into the Hugging Face Hub for sharing. Launched on May 6, 2024 and led by Remi Cadene, a former staff scientist on Tesla's Autopilot and early Optimus humanoid neural networks, LeRobot lets anyone with a laptop and a few hundred dollars of hardware train, fine-tune, and deploy manipulation policies and then publish their datasets and weights for others to reuse ^{[1] [2] [5] [6]}. By May 2026 the Hugging Face Hub hosted more than 58,000 community datasets tagged lerobot, roughly 50 times the count a year earlier, making robotics datasets the single largest dataset category on the Hub ^[24]. By September 16, 2026 the Hub listed 76,464 datasets carrying the LeRobot tag that the library writes into every dataset card it pushes, and 27,610 model repositories tagged with the lerobot library ^{[33] [34]}.

The project's mission is to lower the barrier to entry for robot foundation models so that hobbyist, academic, and industrial robot learning can share one stack. Its 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, WALL-OSS, and NVIDIA's GR00T N1.7; 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 ^[25].

What is LeRobot and who built it?

Origins at Hugging Face

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. Cadene framed the goal directly, describing LeRobot as aiming to be to robotics what the Transformers library is to natural language processing ^[6].

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 project was later documented in an ICLR 2026 paper, posted to arXiv on February 26, 2026, that describes LeRobot as a vertically integrated stack extending from low-level motor middleware through data collection, streaming, policy training, and asynchronous inference ^[26]. The paper tabulates the cost of the supported platforms that publish a bill of materials (roughly 225 euros for an SO-100/101 teleoperation setup and 550 euros for a bimanual pair, about 670 euros for Koch v1.1, about 230 euros for LeKiwi, about 500 euros for the HopeJR arm and hand, and about 21,000 euros for an ALOHA bimanual system) and reports Hub statistics in which the SO-100 and SO-101 accounted for 5,161 and 3,965 openly shared datasets respectively, against 849 for Koch v1.1 and 588 for the Franka Panda, the most downloaded embodiment ^[26]. The same group had earlier published "Robot Learning: A Tutorial" (October 2025), which walks from reinforcement learning and behavior cloning to language-conditioned generalist models using examples implemented in lerobot ^[35].

Why did the SO-100 cheap arm matter?

The single most consequential release of LeRobot's first year was not a model but a piece of hardware. In 2024, Hugging Face partnered with The Robot Studio to publish the SO-100, a 3D-printable arm with a parallel-jaw gripper built around six Feetech STS3215 bus servos ^[7]. The TheRobotStudio/SO-ARM100 repository was created on May 10, 2024 under the Apache 2.0 license, and LeRobot's own support for the arm arrived on October 25, 2024, when Remi Cadene's pull request #419 "Add FeetechMotorsBus, SO-100, Moss-v1" was merged ^{[7] [36]}. 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 kit vendors followed; the repository README as of September 2026 lists frame, electronics, parts and assembled kits from RobotEd, Robonine, PartaBot, ForgeMotion Labs, Seeed Studio, WowRobo, RoboSEasy, NeoBot and Autodiscovery, plus a follower-only SO-100 kit from Phospho ^[7]. An upgraded SO-101 followed on April 28, 2025, with simpler wiring, easier assembly (no gear removal), improved leader-arm motors, and a starting kit price around $100 ^{[8] [9]}. The SO-101 follower keeps six 1/345-geared servos, while the leader mixes three gear ratios (1/191 on the shoulder pan and elbow, 1/345 on the shoulder lift, 1/147 on the wrist and gripper) so that it can hold its own weight yet move with little force ^[8]. The README now treats the SO-101 as the current design and marks the SO-100 documentation, kept in SO100.md, as deprecated; the SO-101's US bill of materials on September 16, 2026 came to $121.94 for one follower and $229.88 for a leader-plus-follower pair, against $123 and $232 in the deprecated SO-100 sheet, and the repository had 7,475 stars that day ^[7]. Naming is loose: the SO-100 CAD assembly files are labelled SO_5DOF_ARM100 (five arm joints plus the gripper), while the SmolVLA paper ("each arm has six degrees of freedom") and NVIDIA's sim-to-real learning path ("a 6-DOF (degrees of freedom) robot arm") count the gripper ^{[7] [61] [62]}. 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] [26]}.

Pollen Robotics acquisition

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.

HopeJR, Reachy Mini, and the humanoid push

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.

How is LeRobot architected?

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 ^{[3] [4]}. 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, shipped with LeRobot 0.4.0 on October 23, 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] [37]}. It remained the current format through the v0.6.1 release: the library's CODEBASE_VERSION constant still reads v3.0, v2.1 datasets must be converted with convert_dataset_v21_to_v30, and the 0.6.1 changelog adds streaming from Hugging Face storage buckets, slice indexing on LeRobotDataset, and a token argument for private Hub datasets rather than a new format ^{[38] [39]}.

LeRobot 0.6.0 release

Hugging Face released LeRobot 0.6.0 on July 7, 2026, following the repository tag on July 6. The release expanded the library from a collection-and-training stack into a more explicit evaluation and correction loop: collect data, train a policy, deploy it, record failures or human interventions, then fine-tune again. Version 0.6.0 remained the latest stable release later that month; v0.6.1 followed on August 3, 2026 ^{[25] [32] [39]}.

Three added policy families use future prediction during training. VLA-JEPA predicts future states in a latent space but removes its world-model branch at inference. LingBot-VA predicts video and actions together and can save imagined video for comparison with the observed rollout. FastWAM pairs a video-generation expert with an action expert during training, then skips video generation when producing actions. The same release added or updated GR00T N1.7, MolmoAct2, EO-1, Multitask DiT, and EVO1 integrations. It also introduced a reward-model interface with Robometer, a 4B model trained from comparisons over more than one million trajectories, and TOPReward, which derives a zero-shot success signal from a vision-language model ^[25].

The data path gained end-to-end depth recording, selectable video codecs, automatic probing for hardware encoders, and timestamped language annotations for subtasks, plans, corrections, speech, and visual questions. The new lerobot-annotate command can populate those fields with a vision-language model. Hugging Face reported up to about twice the previous loading speed from parallel camera decoding and smaller interprocess frame transfers; its episode-subset test fell from 275 seconds to 0.06 seconds. Those figures are project benchmarks rather than independent measurements ^[25].

Evaluation was consolidated under lerobot-eval. Six new integrations, LIBERO-plus, RoboTwin 2.0, RoboCasa365, RoboCerebra, RoboMME, and VLABench, brought the documented total to nine benchmark families when combined with LIBERO, Meta-World, and NVIDIA IsaacLab-Arena. The new lerobot-rollout command became the common path for real-robot deployment. Its DAgger-style strategy lets an operator interrupt a failing policy with a leader arm, records the correction with intervention labels, and returns the resulting dataset to the next fine-tuning pass. LeRobot 0.6.0 also added FSDP training through Accelerate, cloud jobs submitted from lerobot-train, a browser interface called LeLab for SO-ARM101 workflows, and a lighter base installation with feature-specific extras ^[25].

LeRobot 0.6.1 release

Version 0.6.1, tagged on August 3, 2026, was the latest release on GitHub and PyPI as of September 16, 2026, while the main branch carried an unreleased 0.6.2 version string ^{[37] [39] [40]}. Its release notes list one breaking change, the rename of the lerobot.types module to lerobot.lerobot_types, and 146 merged pull requests, most of them fixes ^[39]. The larger structural change was internal: Pi0, Pi0-FAST, Pi0.5, EO-1 and SmolVLA were refactored onto shared VLA components, and the WALL-X and X-VLA ports were rewritten to subclass the native Transformers Qwen2.5-VL and Florence-2 implementations instead of vendoring copies of them ^[39]. User-facing additions included running lerobot-annotate on Hugging Face Jobs, eight robotics-oriented image augmentations, gradient checkpointing for Diffusion Policy, manual exposure, gain and white-balance controls for RealSense cameras, a smooth-handover option for the DAgger and episodic rollout strategies, LeKiwi support in the rollout and evaluation CLIs with a multipart ZMQ observation stream that the notes describe as 25 percent lighter, new Dynamixel motor tables (XH540-W150, XC330-T288, XC330-T181), and a documentation page cataloguing third-party robot and teleoperator plugins ^{[39] [41]}. The release kept the torch>=2.7,<2.12 and Python 3.12 or newer requirements of 0.6.0, so the PyTorch ceiling that the July Jetson field notes had to bypass still applied ^[40].

What policies does LeRobot support?

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 the v0.6.1 source tree, the policies package holds ACT, Diffusion Policy, VQ-BeT, Multitask DiT, TDMPC, Pi0, Pi0-FAST, Pi0.5, SmolVLA, GR00T, X-VLA, WALL-X, EO-1, EVO1, MolmoAct2, VLA-JEPA, LingBot-VA and FastWAM, plus a Real-Time Chunking (RTC) module, added in 0.4.2 (November 2025), that lets flow-matching policies such as Pi0, Pi0.5 and SmolVLA generate the next action chunk while the current one executes ^{[37] [38] [50]}. HIL-SERL lives in the separate rl package as a Soft Actor-Critic learner with human interventions, and the reward models SARM, Robometer and TOPReward have their own documentation section ^{[3] [38]}.

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 June 2025, was particularly important because it was the first VLA the team designed to be cheap enough that a hobbyist could fine-tune it on a single consumer GPU; the team described it as "an open-source, compact, and efficient VLA model that can be trained on consumer-grade hardware using only publicly available datasets" ^[20]. Pretrained on 10 million frames curated from 487 community datasets, the 450M parameter model outperformed much larger VLAs and the ACT baseline on simulation benchmarks (LIBERO, Meta-World) and real SO-100/SO-101 tasks: community pretraining alone raised its success rate on the SO-100 task suite from 51.7 percent to 78.3 percent, a 26.6 percentage point gain ^[20]. Its asynchronous inference path gives roughly 30 percent faster response and twice the task throughput compared to synchronous inference.

What robots does LeRobot support?

LeRobot's hardware support roster has grown from a single Koch v1.1 leader-follower pair, added in July 2024 two months after launch, to thirteen in-tree robot packages by v0.6.1, spanning low-cost desktop arms, bimanual rigs, mobile manipulators, and full humanoids, with a plugin system (introduced in 0.4.0) carrying everything else ^{[37] [38] [41]}. 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. Several can run in either real-world or simulation mode with only a configuration change.

The table lists the robot drivers present in the v0.6.1 source tree (src/lerobot/robots), with the release that introduced each where the release notes record it ^{[37] [38]}.

Beyond the in-tree drivers, the third-party plugin page catalogues packages that LeRobot auto-discovers by the lerobot_robot_ and lerobot_teleoperator_ name prefixes: Trossen's lerobot_trossen for WidowX and ALOHA-style arms, plugins for the UFACTORY xArm, Universal Robots UR5e, Franka, AgileX Piper, ARX5, I2RT YAM and Hiwonder NexArm, VR and haptic teleoperators, and two ROS 2 bridges ^[41]. The roster has also shrunk in places. The v0.3.2 tree contained stretch3 and viperx drivers for the Hello Robot Stretch 3 and Trossen ViperX (ALOHA) arms, and Moss v1 arrived with the SO-100 in 2024; none of those packages appear in the v0.4.0 or later trees, even though the February 2026 paper still lists ALOHA-2 and Stretch 3 among supported platforms ^{[26] [36] [56] [57]}. ALOHA survives in the library as a simulation environment (gym-aloha) and as the origin of the ACT policy ^{[16] [40]}.

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.

What is the LeRobotDataset format?

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 tens of thousands of LeRobotDatasets, growing from roughly 1,145 at the end of 2024 to more than 58,000 by May 2026 and 76,464 with the LeRobot tag on September 16, 2026 ^{[24] [33]}; 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]}.

Which simulators does LeRobot bundle?

LeRobot wraps existing simulation environments in a common gymnasium interface so that policies can be trained, evaluated, and benchmarked without physical hardware. The v0.6.1 package registers pusht (the 2D pushing task introduced as a small Diffusion Policy benchmark, via gym-pusht) and aloha (the bimanual MuJoCo simulation that mirrors the real ALOHA hardware, via gym-aloha) as optional extras, and ships environment modules for LIBERO, Meta-World, RoboCasa, RoboMME, RoboTwin and VLABench ^{[38] [40]}. The documented benchmark families are LIBERO, LIBERO-plus, Meta-World, RoboTwin 2.0, RoboCasa365 (365 kitchen tasks from the RoboCasa project of Nasiriany, Zhu and colleagues, run on a Franka arm mounted on a holonomic base), RoboCerebra, RoboMME (a memory-augmented manipulation benchmark built on ManiSkill and the SAPIEN engine), NVIDIA IsaacLab-Arena, and VLABench; RoboMME, RoboCasa and VLABench are installed outside the normal extras because of dependency pins, and the RoboMME benchmark ships as a Docker image ^{[38] [40] [58]}. LIBERO and Meta-World are the two benchmarks SmolVLA's evaluation table reports ^[20]. EnvHub, added in 0.4.1 (November 2025), loads environments published as Git repositories on the Hub, and the LeIsaac package uses it to teleoperate an SO-101 in Isaac Lab, record demonstrations and train in LeRobot ^{[37] [59]}. 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 model without ever touching physical hardware ^[21].

How big is the LeRobot community?

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. On September 16, 2026 the huggingface/lerobot repository had 27,527 stars and 5,658 forks, and the TheRobotStudio/SO-ARM100 hardware repository 7,475 stars and 685 forks ^{[1] [7]}. The clearest measure of momentum is the Hub itself: LeRobot-tagged datasets grew roughly 50-fold in a single year to more than 58,000 by May 2026, making robotics the single largest dataset category on the Hugging Face Hub, a milestone practitioners cited as the point where open-source robot learning crossed from research infrastructure into production-grade tooling ^[24]. 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.

Notable collaborations

Physical Intelligence

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

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.

July 2026 NVIDIA integrations

LeRobot 0.6.0 replaced its GR00T N1.5 port with GR00T N1.7. Hugging Face and NVIDIA said the port was parity-tested against the original Isaac-GR00T implementation for matching inputs and outputs; users who still need N1.5 must pin LeRobot 0.5.1. A companion Isaac Teleop workflow connects an XR controller to an SO-101 over CloudXR/OpenXR, records demonstrations into a LeRobotDataset, post-trains GR00T N1.7, and deploys the result with lerobot-rollout. The published training example used an RTX 6000 Pro, so it does not establish that GR00T post-training fits on a Jetson module ^{[25] [27]}.

On July 22, NVIDIA Robotics promoted a separate tutorial by Iulia Feroli that described going from a fresh Jetson installation to leader-follower SO-101 teleoperation in about 30 minutes. The linked material is an edited video running 8 minutes 50 seconds on the Back to Engineering channel, made with Seeed Studio, plus a public setup repository. It is a community procedure amplified by NVIDIA, not an NVIDIA-authored compatibility guide or a timed independent benchmark ^{[28] [29] [30]}.

The tested machine was a Seeed reComputer Super with a Jetson Orin NX 16GB. Its stack used JetPack 7.2 and Jetson Linux 39.2, Ubuntu 24.04, CUDA 13.2, Python 3.12, ROS 2 Jazzy, prerelease torch 2.12.1+cu132 and torchvision 0.27.1+cu132 wheels, NumPy 2.2.6, and LeRobot 0.6.0. NVIDIA's release page independently confirms that JetPack 7.2 added Orin-family support and ships Jetson Linux 39.2, Ubuntu 24.04, and CUDA 13.2.1 ^{[30] [31]}.

The procedure flashes the board, exposes CUDA through environment variables, creates a Python 3.12 environment, verifies a real CUDA matrix multiplication, installs ROS 2 Jazzy in the system environment, and installs LeRobot and the Feetech motor packages before assigning motor IDs, calibrating both arms, and starting teleoperation. The separation matters because the guide deliberately keeps ROS and the PyTorch environment apart. Hugging Face's standard LeRobot documentation supports Python 3.12 or newer, but its normal Linux installation path does not document this JetPack-specific cu132 combination ^{[30] [32]}.

The guide also records unresolved version coupling. Its PyTorch wheels were prerelease builds outside NVIDIA's published compatibility matrix, and LeRobot 0.6.0's declared PyTorch ceiling excluded the tested 2.12.1 build. To preserve the Orin-compatible CUDA wheel, the procedure installs LeRobot without dependency resolution and adds required packages manually, leaving expected pip check incompatibilities. The repository says the path was tested only on its Orin NX 16GB setup and labels the instructions as field notes. The demonstration proves local SO-101 teleoperation on that configuration; it does not show autonomous policy inference, policy training on Jetson, or ROS-native LeRobot control ^{[30] [32]}.

TheRobotStudio

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.

Reception and impact

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 (ROBOTIS and Seeed Studio in the core library, Trossen through its own plugin, and Hello Robot in the 2024-2025 codebase), while others have positioned themselves as offering closed alternatives ^{[38] [41] [56]}. 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.

Is LeRobot finished, and what are its limits?

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.

Embedded deployment compatibility: LeRobot depends on Python, PyTorch, video, camera, and motor packages that do not always publish matching builds for each embedded GPU and operating-system release. The July 2026 JetPack 7.2 SO-101 procedure used prerelease CUDA wheels and bypassed LeRobot's declared PyTorch ceiling, which v0.6.1 left unchanged at torch<2.12 ^[40]. A working teleoperation demo on one Orin NX configuration is therefore not a general support guarantee for other Jetson boards or for training and autonomous inference workloads ^{[30] [32]}.

See also

• Rerun (rerun.io)
• SO-101
• OpenArm
• reBot Arm
• Unitree G1
• Hugging Face
• Robotics
• Robot foundation model
• Vision-language-action model
• Imitation learning
• Behavior cloning
• NVIDIA Isaac Sim
• PyTorch
• Diffusion models
• Reinforcement learning

References

1. ^1 ^2 ^3 ^4Hugging Face. *LeRobot GitHub repository*. github.com/...lerobot
2. ^Cadene, R. *Personal site*. remicadene.com
3. ^1 ^2Hugging Face. *LeRobot documentation*. huggingface.co/...index
4. ^Hugging Face. *lerobot Hub organization*. huggingface.co/lerobot
5. ^1 ^2VentureBeat. (2024, March 7). *Hugging Face is launching an open source robotics project led by former Tesla scientist*. venturebeat.com/...t-led-by-former-tesla-scientist
6. ^1 ^2 ^3 ^4VentureBeat. (2024, May 6). *Hugging Face launches LeRobot open source robotics code library*. venturebeat.com/...en-source-robotics-code-library
7. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8TheRobotStudio. *SO-ARM100 GitHub repository*. github.com/...SO-ARM100
8. ^1 ^2 ^3Hugging Face. *SO-101 documentation*. huggingface.co/...so101
9. ^1 ^2TechCrunch. (2025, April 28). *Hugging Face releases a 3D-printed robotic arm starting at $100*. techcrunch.com/...nted-robotic-arm-starting-at-100
10. ^Hugging Face. *SO-100 documentation*. huggingface.co/...so100
11. ^1 ^2 ^3Hugging Face. (2025, April 14). *Hugging Face to sell open-source robots thanks to Pollen Robotics acquisition*. huggingface.co/...face-pollen-robotics-acquisition
12. ^Fortune. (2025, April 14). *AI company Hugging Face buys humanoid robot company Pollen Robotics, maker of Reachy 2*. fortune.com/...ot-company-pollen-robotics-reachy-2
13. ^1 ^2 ^3TechCrunch. (2025, May 29). *Hugging Face unveils two new humanoid robots*. techcrunch.com/...-unveils-two-new-humanoid-robots
14. ^Hugging Face. *HopeJR documentation*. huggingface.co/...hope_jr
15. ^1 ^2 ^3 ^4Hugging Face. *LeRobotDataset v3.0: Bringing large-scale datasets to lerobot*. huggingface.co/...lerobot-datasets-v3
16. ^1 ^2Zhao, T., Kumar, V., Levine, S., and Finn, C. (2023). *Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware*. arXiv:2304.13705. tonyzhaozh.github.io/aloha
17. ^Chi, C. et al. (2023). *Diffusion Policy: Visuomotor Policy Learning via Action Diffusion*. RSS 2023. diffusion-policy.cs.columbia.edu
18. ^1 ^2Physical Intelligence. (2024). *Pi0: A Vision-Language-Action Flow Model for General Robot Control*. pi.website/...pi0.pdf
19. ^1 ^2Physical Intelligence. (2025). *Pi0.5: A Vision-Language-Action Model with Open-World Generalization*. arXiv:2504.16054. arxiv.org/...2504.16054
20. ^1 ^2 ^3 ^4 ^5Hugging Face. (2025, June). *SmolVLA: Efficient Vision-Language-Action Model trained on LeRobot Community Data*. huggingface.co/...smolvla
21. ^1 ^2 ^3 ^4NVIDIA. (2025, March 18). *NVIDIA Announces Isaac GR00T N1 the World's First Open Humanoid Robot Foundation Model*. nvidianews.nvidia.com/...del-simulation-frameworks
22. ^NVIDIA. *Isaac-GR00T GitHub repository*. github.com/...Isaac-GR00T
23. ^Grokipedia. *LeRobot*. grokipedia.com/...LeRobot
24. ^1 ^2 ^3TechTimes. (2026, May 25). *Open Source Robotics AI Reaches Inflection Point: LeRobot Hub Surpasses 58,000 Datasets in One Year*. techtimes.com/...surpasses-58000-datasets-one-year
25. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10Hugging Face. (2026, July 7). *LeRobot v0.6.0: Imagine, Evaluate, Improve*. huggingface.co/...lerobot-release-v060
26. ^1 ^2 ^3 ^4 ^5 ^6 ^7Cadene, R. et al. (2026). *LeRobot: An Open-Source Library for End-to-End Robot Learning*. ICLR 2026 (poster); arXiv:2602.22818. arxiv.org/...2602.22818 ; OpenReview: openreview.net/forum
27. ^1 ^2NVIDIA and Hugging Face. (2026, July 7). *NVIDIA Isaac Teleop and GR00T 1.7 Open VLA Model Available in LeRobot*. huggingface.co/...ac-teleop-and-gr00t17-in-lerobot
28. ^NVIDIA Robotics. (2026, July 22). *From a fresh install to teleoperating SO-101 robot arms in only 30 minutes*. X. x.com/...2079959655409717284
29. ^Back to Engineering. (2026, July 15). *NVIDIA JetPack 7.2 Full Setup: PyTorch, ROS 2 & LeRobot in 30 Minutes (2026)*. YouTube. youtube.com/watch
30. ^1 ^2 ^3 ^4 ^5Feroli, I. (2026). *PyTorch on JetPack 7.2 for Jetson Orin NX (the undocumented path)*. GitHub. github.com/...jetson-jp7.2-install
31. ^NVIDIA. (2026). *JetPack SDK Downloads and Notes: JetPack 7.2 with Jetson Linux 39.2*. developer.nvidia.com/...downloads
32. ^1 ^2 ^3 ^4Hugging Face. *LeRobot installation*. huggingface.co/...installation
33. ^1 ^2Hugging Face Hub. *Datasets tagged LeRobot* (76,464 listed on September 16, 2026). huggingface.co/datasets
34. ^Hugging Face Hub. *Models using the lerobot library* (27,610 listed on September 16, 2026). huggingface.co/models
35. ^Capuano, F., Pascal, C., Zouitine, A., Wolf, T., and Aractingi, M. (2025, October 14). *Robot Learning: A Tutorial*. arXiv:2510.12403. arxiv.org/...2510.12403
36. ^1 ^2 ^3Cadene, R. (2024, October 25). *Add FeetechMotorsBus, SO-100, Moss-v1*. huggingface/lerobot pull request #419. github.com/...419
37. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14Hugging Face. *LeRobot releases* (v0.3.2 through v0.6.1, with per-release changelogs). github.com/...releases
38. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10Hugging Face. *LeRobot source tree at v0.6.1* (policies, robots, teleoperators, envs, datasets packages). github.com/...lerobot
39. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8Hugging Face. (2026, August 3). *Release v0.6.1*. github.com/...v0.6.1
40. ^1 ^2 ^3 ^4 ^5 ^6Hugging Face. *pyproject.toml at v0.6.1* (version, Python and PyTorch requirements, optional extras). github.com/...pyproject.toml
41. ^1 ^2 ^3 ^4Hugging Face. *Third-Party Robots & Teleoperators*. LeRobot documentation. huggingface.co/...third_party_robots
42. ^Hugging Face. *X-VLA*. LeRobot documentation. huggingface.co/...xvla
43. ^Hugging Face. *WALL-OSS*. LeRobot documentation. huggingface.co/...walloss
44. ^Hugging Face. *EO-1*. LeRobot documentation. huggingface.co/...eo1
45. ^Hugging Face. *EVO1*. LeRobot documentation. huggingface.co/...evo1
46. ^Hugging Face. *MolmoAct2 Policy*. LeRobot documentation. huggingface.co/...molmoact2
47. ^Hugging Face. *VLA-JEPA*. LeRobot documentation. huggingface.co/...vla_jepa
48. ^Hugging Face. *LingBot-VA*. LeRobot documentation. huggingface.co/...lingbot_va
49. ^Hugging Face. *FastWAM*. LeRobot documentation. huggingface.co/...fastwam
50. ^Hugging Face. *Real-Time Chunking (RTC)*. LeRobot documentation. huggingface.co/...rtc
51. ^Hugging Face. *Unitree G1*. LeRobot documentation. huggingface.co/...unitree_g1
52. ^Hugging Face. *EarthRover Mini Plus*. LeRobot documentation. huggingface.co/...earthrover_mini_plus
53. ^Hugging Face. *OMX*. LeRobot documentation. huggingface.co/...omx
54. ^Hugging Face. *OpenArm*. LeRobot documentation. huggingface.co/...openarm
55. ^Hugging Face. *reBot B601-DM*. LeRobot documentation. huggingface.co/...rebot_b601
56. ^1 ^2Hugging Face. *LeRobot source tree at v0.3.2, robots package* (includes `stretch3` and `viperx`). github.com/...robots
57. ^Hugging Face. *LeRobot source tree at v0.4.0, robots package*. github.com/...robots
58. ^Hugging Face. *RoboMME*. LeRobot documentation. huggingface.co/...robomme
59. ^Hugging Face. *LeIsaac x LeRobot EnvHub*. LeRobot documentation. huggingface.co/...envhub_leisaac
60. ^Kooijmans, P. (2025, July 16). *Feat/add bimanual so100 robot*. huggingface/lerobot pull request #1509. github.com/...1509
61. ^Shukor, M., Aubakirova, D., Capuano, F., et al. (2025, June 2). *SmolVLA: A Vision-Language-Action Model for Affordable and Efficient Robotics*. arXiv:2506.01844. arxiv.org/...2506.01844
62. ^NVIDIA. *Train an SO-101 Robot From Sim-to-Real With NVIDIA Isaac: LeRobot: Background and Community*. docs.nvidia.com/...04-lerobot