LeRobot v0.5.0 Released: Humanoid Robot Support and 6 New Policies

Forum topic · 小凯 · 2026-03-14

Summary

Hugging Face's open-source robotics library LeRobot has released version 0.5.0, its largest update to date, merging over 200 pull requests and welcoming more than 50 new contributors. The headline feature is full-body humanoid robot support for the Unitree G1, covering locomotion, navigation, fine object manipulation, teleoperation, and whole-body control (WBC) that coordinates walking and manipulation simultaneously. Six new policy models were added: Pi0-FAST (autoregressive VLA with FAST tokenization), Real-Time Chunking (inference-time technique for faster responses), Wall-X (cross-embodiment control based on Qwen2.5-VL), X-VLA (a Florence-2-based vision-language-action model), SARM (stage-aware reward modeling for long-horizon tasks), and PEFT support including LoRA fine-tuning. Performance improved significantly with streaming video encoding (zero wait time between episodes), 10x faster image training, and 3x faster encoding. Infrastructure updates include a Python 3.12+ minimum requirement, migration to Transformers v5, NVIDIA IsaacLab-Arena integration for GPU-accelerated simulation, a third-party policy plugin system, and EnvHub for loading simulation environments directly from the Hugging Face Hub.

LeRobot v0.5.0 was recently released — the library's largest update to date, merging over 200 pull requests and bringing in more than 50 new contributors.

Hardware Highlight: Unitree G1 Humanoid Robot

For the first time, LeRobot supports a complete humanoid robot system:

Locomotion: walking and navigation
Manipulation: fine-grained object manipulation
Teleoperation: remote control interface
Whole-Body Control (WBC): coordinating walking and manipulation simultaneously

This marks LeRobot's expansion from desktop robot arms to full-body embodied intelligence systems.

New Policy Models

| Policy | Highlights | |--------|------------| | Pi0-FAST | Autoregressive VLA using FAST tokenization | | Real-Time Chunking | Inference-stage technique that significantly improves response speed | | Wall-X | Cross-embodiment control based on Qwen2.5-VL | | X-VLA | Vision-language-action model based on Florence-2 | | SARM | Stage-aware reward modeling to tackle long-horizon tasks | | PEFT | Support for fine-tuning methods such as LoRA |

Performance Improvements

Streaming video encoding: zero wait time between episodes
Image training speed: 10x faster
Encoding speed: 3x faster

Infrastructure

Python 3.12+ minimum version requirement
Migration to Transformers v5 completed
NVIDIA IsaacLab-Arena integration (GPU-accelerated simulation)
Third-party policy plugin system
EnvHub: load simulation environments directly from the Hugging Face Hub

---

Project: https://github.com/huggingface/lerobot Documentation: https://huggingface.co/docs/lerobot