What Hugging Face released

On March 9, 2026, Hugging Face released LeRobot v0.5.0, a major update to its open robot-learning stack. The company describes it as the project’s biggest release yet, with more than 200 merged pull requests and more than 50 new contributors since v0.4.0. The important point is not just scale. This release expands hardware support, policy options, dataset handling, simulation distribution, and the core codebase all at once.

The biggest single addition is full support for the Unitree G1. Hugging Face says this is LeRobot’s first humanoid integration and that it includes locomotion, manipulation, teleoperation, and whole-body control. That pushes the project beyond tabletop arms toward full-body embodied AI workflows.

What is new in v0.5.0

Hardware expansion: besides Unitree G1, the release adds OpenArm, OpenArm Mini, Earth Rover, OMX Robot, and new CAN-bus motor-controller support.
Policy expansion: new options include Pi0-FAST with a Gemma 300M-based autoregressive action expert, Real-Time Chunking for more responsive inference, Wall-X, X-VLA, SARM, and LoRA/PEFT fine-tuning support.
Dataset performance: streaming video encoding removes post-episode wait time, while Hugging Face says image training is now up to 10x faster and encoding up to 3x faster.
Simulation and sharing: the release introduces EnvHub, which loads environments directly from the Hugging Face Hub, and adds NVIDIA IsaacLab-Arena integration.

The codebase itself also moves forward. LeRobot now targets Python 3.12+ and Transformers v5, adds third-party policy plugins, improves remote visualization, and updates PyTorch version bounds to support NVIDIA Blackwell GPUs. That makes the release look less like a single robotics repo update and more like a platform maturing around a community and tooling ecosystem.

Why it matters

Open robotics is constrained by much more than model quality. Teams need consistent ways to connect hardware, datasets, simulators, training code, and deployment workflows. LeRobot v0.5.0 is notable because it tries to improve all of those layers together rather than shipping one isolated model result.

The first humanoid integration is especially significant. As embodied AI work shifts toward general-purpose humanoids and longer-horizon manipulation, open stacks that can handle real hardware, simulation, and policy iteration quickly become strategic infrastructure. Hugging Face is positioning LeRobot as that kind of infrastructure, not just as a collection of robotics demos.

Source: Hugging Face Blog

#lerobot

Hugging Face expands LeRobot with first humanoid support, faster datasets, and a broader robot-learning stack