Hugging Face ships LeRobot 0.6 with world models and NVIDIA's GR00T N1.7, deepening an open robotics alliance

Hugging Face has turned its LeRobot library into a shelf of competing robot brains. Version 0.6.0, released on July 7, 2026, adds three world-model policies, five vision-language-action (VLA) models, two reward models and six simulation benchmarks, and it arrives with a package of NVIDIA tools that the chipmaker announced for LeRobot the day before.

The NVIDIA side of the release has five parts. Isaac GR00T N1.7, which NVIDIA calls the first open and commercially viable robot foundation model, can now be post-trained and deployed through LeRobot workflows. Isaac Teleop, an open-source data-collection framework, captures human demonstrations from external devices in standardised formats directly in LeRobot. Isaac Lab-Arena joins the LeRobot Environment Hub so developers can build simulation scenes and use them to train and evaluate policies such as GR00T, Physical Intelligence's Pi models and SmolVLA. Jetson Thor support targets deployment of VLA models on LeRobot's Reachy 2 open-source humanoid. NVIDIA said its Cosmos 3 world model is coming soon to LeRobot.

NVIDIA framed the partnership in audience terms: it connects NVIDIA's 3 million robotics developers with Hugging Face's 16 million AI builders. It also pointed to what it calls the largest open-source physical AI dataset, downloaded more than 15 million times, with more than 350,000 real and simulated trajectories and 57 million grasps. Hugging Face co-founder and chief science officer Thomas Wolf said open source is how a field turns advanced research into something people can study, adapt and build on.

The model list in LeRobot 0.6 is broad. The three world-model policies are VLA-JEPA, a compact model built on Qwen3-VL-2B that predicts the future in latent space; LingBot-VA, an autoregressive video-action model that needs a 24 to 32 GB GPU; and FastWAM, a model of about 5 billion parameters that pairs video-generation and action experts. The five VLAs are GR00T N1.7, which uses NVIDIA's Cosmos-Reason2-2B vision-language model; MolmoAct2 from the Allen Institute for AI, which needs about 12 GB for inference and can run zero-shot on SO-100 and SO-101 arms; EO-1 on a Qwen2.5-VL-3B backbone; Multitask DiT, a diffusion transformer of about 450 million parameters; and EVO1, a 0.77-billion-parameter model on InternVL3-1B.

The two reward models address a bottleneck in robot learning, judging whether an attempt worked. Robometer is a general-purpose reward model built on Qwen3-VL-4B and trained on more than 1 million robot trajectories. TOPReward scores attempts zero-shot by reading the probability that a vision-language model answers "True".

The six new benchmarks, all run through the lerobot-eval command, bring the library to nine benchmark families including LIBERO, Meta-World and NVIDIA's IsaacLab-Arena. They include LIBERO-plus with 10,000 perturbed task variants, RoboTwin 2.0 with 50 bimanual tasks and more than 100,000 training trajectories, RoboCasa365 with 365 kitchen tasks across 2,500 procedurally generated kitchens, RoboCerebra with a 6,660-episode long-horizon dataset, RoboMME with 16 memory tasks, and VLABench for knowledge and reasoning.

Under the hood, the release adds fully sharded data-parallel training across GPUs, one-flag cloud training on Hugging Face Jobs, data loading up to twice as fast, a lerobot-rollout command with human-in-the-loop corrections, depth-camera support for Intel RealSense and an automatic language-annotation pipeline. Loading a dataset subset dropped from 275 seconds to 0.06 seconds, and base dependencies shrank by 40%. Supported hardware includes the SO-100 and SO-101 arms, Koch, OpenArm, reBot and Unitree's G1 humanoid. The release lists 10 core authors and more than 79 community contributors.

GR00T N1.7 itself dates from April 17, when NVIDIA opened early access and added a licence it says supports production use. Its Hugging Face model card lists 3 billion parameters, training on 21.6 million data points from 13 datasets mixing human, robot and simulated data, and release under the NVIDIA Open Model License. The model card reports up to 35.9 Hz inference on an H100 GPU with TensorRT.

The competitive context explains why NVIDIA is investing in someone else's library. Google DeepMind's Gemini Robotics 2, released on July 30, keeps its action models with trusted testers. Skild AI, which introduced its S1 model on August 18, sells a proprietary brain. Physical Intelligence open-sourced its π0 policy in February 2025, and its models are among those LeRobot's environment hub supports. By making LeRobot the place where open policies are trained and compared, NVIDIA puts GR00T, Isaac and Jetson next to every rival model a developer tries.

The risks are fragmentation and quality. Eight new policies in one release give developers choice but little guidance on which works for a given robot, and the benchmarks are simulations rather than factory tests. Hugging Face does not publish usage figures for individual policies, and NVIDIA has not given a date for the Cosmos 3 integration.