LeRobot v0.5.0: A Revolution in Robotics with the Unitree G1

LeRobot v0.5.0: A Major Update

Version v0.5.0 of LeRobot marks a significant advancement with over 200 PRs merged and the addition of 50 new contributors since version v0.4.0. This update, the most important to date, expands in all directions, offering improvements in both hardware and software. Among the new features are the addition of robots, including the first humanoid, advanced policies, and a modernized codebase running on Python 3.12 and Transformers v5.

The return of autoregressive VLAs is also a highlight of this version, featuring faster datasets and simulation environments that can now be loaded directly from the Hub. The integration of NVIDIA IsaacLab-Arena and a streaming video encoding that eliminates wait times between episode recordings further enrich the user experience.

The Unitree G1: A Leap Forward for Humanoids

LeRobot v0.5.0 introduces full support for the Unitree G1 humanoid, marking the first integration of a humanoid robot into the ecosystem. This robot is capable of locomotion, delicate object manipulation, and teleoperation via an intuitive interface. Full Body Control (WBC) allows for the coordination of locomotion and manipulation for complex tasks, representing a step towards general-purpose robotics.

The Unitree G1 can walk, navigate, and move in various environments while performing delicate object manipulation tasks. Teleoperation enables remote control of the G1 with an intuitive interface, and Full Body Control (WBC) ensures simultaneous coordination of locomotion and manipulation for complex real-world tasks.

New Hardware Additions

In addition to the Unitree G1, LeRobot v0.5.0 expands its hardware support with the OpenArm robot and its Mini model, allowing for bi-manual configurations for complex manipulation tasks. The integration of the Earth Rover now enables outdoor navigation, while the OMX Robot offers configurable grip settings. The consolidation of SO-100 and SO-101 implementations simplifies maintenance while maintaining performance.

A new motor controller support via the CAN (Controller Area Network) bus paves the way for high-performance actuators. The new motor controllers, such as RobStride and Damiao, allow for a wider range of hardware to be operated, thereby enhancing the capabilities of the LeRobot ecosystem.

Advances in Policies and Simulation

This version also enriches the policy offerings with the introduction of six new techniques, including Pi0-FAST. This autoregressive Vision-Language-Action model utilizes FAST tokenization for flexible and responsive decoding. Pi0-FAST, based on the autoregressive action expert Gemma 300M, generates discretized action tokens, enabling efficient autoregressive tokenization and decoding.

Furthermore, the integration of EnvHub allows for the loading of simulation environments directly from the Hugging Face Hub, facilitating the development and testing of new robotic applications. Real-time chunking for responsive inference is also supported, providing an enhanced user experience. With these advancements, LeRobot v0.5.0 paves the way for more varied and high-performing robotic applications, strengthening its position in the field of open-source robotic learning.