← Back
Hugging Face
LeRobot v0.5.0 adds humanoid support and autoregressive VLAs with 3x faster video encoding
· releasefeaturemodelplatformintegrationperformance · huggingface.co ↗

Robots Expanded

LeRobot v0.5.0 introduces Unitree G1 humanoid support, marking the framework's first full humanoid integration. The G1 implementation includes locomotion, dexterous manipulation, teleoperation, and whole-body control (WBC) for coordinated arm-and-body tasks. Beyond the humanoid, the release adds OpenArm and OpenArm Mini robot integrations, CAN bus motor support, and numerous other hardware additions — significantly expanding the ecosystem from tabletop arms to mobile and full-body systems.

Policy & Inference Improvements

The release debuts Pi0-FAST, an autoregressive Vision Language Action (VLA) model for responsive robot control. Real-Time Chunking (RTC) enables low-latency streaming inference, allowing robots to act on visual input with minimal delay. Additional policies added include Wall-X, X-VLA, and SARM, alongside PEFT support for parameter-efficient fine-tuning — giving developers multiple approaches to policy learning and deployment.

Performance Gains

Streaming video encoding eliminates idle time between recording episodes, with 3x faster overall encoding speed. Image training is 10x faster with optimized data pipelines. These improvements dramatically reduce dataset preparation overhead, enabling faster iteration on robotics research and deployment.

Simulation & Infrastructure

EnvHub introduces a new capability for loading simulation environments directly from Hugging Face Hub, complemented by NVIDIA IsaacLab-Arena integration. The codebase has been modernized to Python 3.12 and Transformers v5, with plugin support for third-party policies. Over 200 PRs from 50+ new contributors shaped this release, reflecting strong community momentum.

What Developers Should Know

  • Training and deployment pipelines now support humanoid robots and hybrid arm-manipulation tasks
  • Streaming encoding and faster training pipelines reduce iteration time significantly
  • New policy options (autoregressive, chunking-based, fine-tuning) allow flexible algorithm selection
  • EnvHub integration simplifies simulation setup and experiment reproducibility
  • Python 3.12 and Transformers v5 requirements mean dependency updates are necessary