LeRobot v0.5.0 release notes / what's new

LeRobot v0.5.0 is a major expansion of the Hugging Face robotics library, introducing support for humanoid hardware, a wider variety of Vision-Language-Action (VLA) policies, and significant performance gains in data recording and training. This release focuses on scaling the ecosystem across hardware, models, and data pipelines.

Hardware Expansion and Humanoid Support

LeRobot v0.5.0 introduces comprehensive support for the Unitree G1 humanoid, marking the library's first humanoid integration. This support includes whole-body control (WBC) to coordinate locomotion and manipulation simultaneously, as well as dedicated interfaces for teleoperation and navigation.

Other hardware additions include:

  • OpenArm & OpenArm Mini: Support for the OpenArm robot arm and its companion teleoperator, including bi-manual configurations.
  • Earth Rover: The first mobile robot integration for outdoor navigation.
  • OMX Robot: A new robot arm featuring calibration support and configurable gripper settings.
  • CAN Bus Motor Support: Integration for RobStride and Damiao CAN-based motor controllers, enabling the use of professional-grade, high-torque actuators.
  • SO-100/SO-101: Implementation consolidation into a single codebase for easier maintenance of bi-manual setups.

New Policies and Model Zoo

This release adds six new policies and techniques to enhance robot learning and inference responsiveness.

Vision-Language-Action (VLA) Models

  • Pi0-FAST: An autoregressive VLA using Frequency-space Action Sequence Tokenization (FAST). It utilizes an action expert based on Gemma 300M to generate discretized action tokens.
  • Wall-X: A VLA policy based on Qwen2.5-VL that employs a flow-matching head for cross-embodiment robotic control.
  • X-VLA: A Florence2-based VLA providing an alternative backbone for robot learning.

Inference and Training Enhancements

  • Real-Time Chunking (RTC): An inference-time technique that blends new predictions with in-progress actions. This reduces latency and makes flow-matching policies (such as the Pi0 family, SmolVLA, and Diffusion) more responsive.
  • SARM (Stage-Aware Reward Modeling): A method for long-horizon tasks that predicts both the task stage and progress within that stage, rather than relying on a single global linear progress signal.
  • PEFT Support: Large VLAs can now be fine-tuned using LoRA and other Parameter-Efficient Fine-Tuning methods via policy-level configuration.

Dataset Performance and Tooling

LeRobot v0.5.0 optimizes the data pipeline to eliminate bottlenecks in recording and training.

Recording and Encoding

  • Streaming Video Encoding: Frames are now encoded in real-time during capture, eliminating the wait time between recording episodes. This feature includes automatic hardware encoder detection for GPU acceleration.
  • Encoding Speed: Parallel encoding is now the default, resulting in 3x faster encoding overall.

Training and Management

  • Training Speed: Image training is now 10x faster due to the removal of data access bottlenecks and improved image transform support.
  • Dataset Tools: New capabilities include subtask annotation for hierarchical learning, image-to-video conversion for storage efficiency, and an info operation for dataset inspection.

EnvHub and Simulation Integration

EnvHub allows users to load simulation environments directly from the Hugging Face Hub using HubEnvConfig. This removes the need for local package installation by downloading and executing remote make_env functions.

Additionally, the release integrates NVIDIA IsaacLab-Arena, providing GPU-accelerated, massively parallel environment instances for reinforcement learning based on NVIDIA Isaac Sim.

Codebase Modernization

LeRobot v0.5.0 updates its technical foundation to support modern hardware and software standards:

  • Core Requirements: Minimum requirement updated to Python 3.12+ and migration to Transformers v5.
  • Extensibility: Support for third-party policy plugins allows users to register custom policies as installable packages via pip without modifying the core library.
  • Hardware Support: PyTorch version bounds have been updated to support NVIDIA Blackwell GPUs.
  • Telemetry: Remote Rerun visualization now supports compressed images for efficient bandwidth usage.

Sources