Hardware Expansion: Humanoid Support
LeRobot v0.5.0 introduces its first humanoid integration with full Unitree G1 support. This comprehensive integration includes locomotion, dexterous manipulation, teleoperation, and whole-body control (WBC) that coordinates walking and object manipulation simultaneously. The release also adds support for OpenArm & OpenArm Mini robots and CAN bus motor integration, expanding the platform's hardware ecosystem beyond tabletop arms.
New Policy Architectures and Inference Methods
The release features several new policy options for developers:
- Pi0-FAST: Autoregressive vision language action models making a comeback for responsive inference
- Real-Time Chunking (RTC): New inference method enabling lower-latency robot control
- Wall-X, X-VLA, and SARM: Additional policy variants now available
- PEFT Support: Parameter-efficient fine-tuning integration for faster model adaptation
Performance Improvements
Significant performance gains help accelerate development workflows:
- Streaming Video Encoding: Eliminates idle time between recording episodes by encoding video on-the-fly
- 10x faster image training with optimized data loading
- 3x faster encoding compared to previous versions
- EnvHub integration allows loading simulation environments directly from Hugging Face Hub
Codebase Modernization
The v0.5.0 release includes major infrastructure upgrades running on Python 3.12+, Transformers v5, and support for third-party policy plugins. Over 200 merged PRs and 50 new contributors shaped this release, representing the largest update to the LeRobot ecosystem to date. NVIDIA IsaacLab-Arena integration provides simulation capabilities for policy development and testing.