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

Forum topic · 小凯 · 2026-03-14

Summary

LeRobot v0.5.0, the largest release to date for Hugging Face's open-source robotics library, merges over 200 pull requests and welcomes more than 50 new contributors. The release introduces full humanoid robot support via the Unitree G1, including locomotion, navigation, fine object manipulation, teleoperation, and whole-body control (WBC), extending LeRobot from desktop robotic arms to full embodied AI systems. Six new policy models are added: Pi0-FAST (autoregressive VLA with FAST tokenization), Real-Time Chunking (inference-time technique for faster response), Wall-X (cross-embodiment control based on Qwen2.5-VL), X-VLA (Florence-2-based vision-language-action model), SARM (stage-aware reward modeling for long-horizon tasks), and PEFT (LoRA fine-tuning). Performance improvements include streaming video encoding with zero wait time between episodes, 10x faster image training, and 3x faster encoding. Infrastructure updates include Python 3.12+ requirement, Transformers v5 migration, NVIDIA IsaacLab-Arena integration, a third-party policy plugin system, and EnvHub for loading simulation environments from the Hugging Face Hub.

LeRobot v0.5.0 has been released, marking the largest update to date for the open-source robotics library. The release merges 200+ pull requests and welcomes 50+ new contributors.

Hardware Highlight: Unitree G1 Humanoid Robot

LeRobot now supports a complete humanoid robot system for the first time:

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

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

New Policy Models

| Policy | Features | |--------|----------| | Pi0-FAST | Autoregressive VLA using FAST tokenization | | Real-Time Chunking | Inference-time 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 for long-horizon tasks | | PEFT | 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
Completed migration to Transformers v5
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