Major Hardware Expansion
LeRobot v0.5.0 introduces full Unitree G1 humanoid support, marking the framework's first humanoid integration. The G1 support includes locomotion, dexterous manipulation, teleoperation, and whole-body control (WBC) to coordinate movement and arm control simultaneously. Beyond the humanoid, the release adds support for additional arms including OpenArm and OpenArm Mini, CAN Bus motor integration, and more mobile robots.
New Policies and Model Architectures
The release brings significant expansions to the policy zoo:
- Pi0-FAST: Autoregressive vision-language action models making a comeback to LeRobot
- Real-Time Chunking (RTC): New inference technique for responsive, real-time policy execution
- Wall-X, X-VLA, SARM: Additional policy architectures now supported
- PEFT Integration: Parameter-efficient fine-tuning support for adapting policies with minimal compute
Performance and Workflow Improvements
Dataset handling receives major improvements:
- Streaming Video Encoding: Eliminates wait times between recording episodes by encoding video on-the-fly
- 10x Faster Image Training: Substantial speedup for training on image-based policies
- 3x Faster Encoding: Improved video compression pipeline
- EnvHub: Load simulation environments directly from the Hugging Face Hub alongside NVIDIA IsaacLab-Arena integration
Modernized Codebase
The release upgrades the foundation with Python 3.12+ and Transformers v5, enabling better performance and access to latest model architectures. Third-party policy plugins are now supported, allowing community contributions without modifying core code.
Release Scale
With over 200 merged PRs and 50+ new contributors since v0.4.0, this represents LeRobot's largest release to date, demonstrating strong ecosystem growth and community engagement.