LeRobot v0.5.0: Scaling Every Dimension

Open-source robotics tooling doesn't usually move this fast, but LeRobot v0.5.0 is a reminder that Hugging Face's robot-learning project is scaling in every direction at once. Over 200 merged pull requests and more than 50 new contributors since v0.4.0 landed a release that touches hardware, models, datasets, and the codebase itself. The headline item is support for the Unitree G1, LeRobot's first humanoid, complete with locomotion, manipulation, teleoperation, and whole-body control that lets a robot walk and use its arms at the same time.

The hardware list doesn't stop there. LeRobot picked up the OpenArm and OpenArm Mini teleoperator, added its first mobile robot with Earth Rover, folded in the OMX arm, and cleaned up years of duplicated code by merging the SO-100 and SO-101 implementations into one shared base. New CAN bus support for RobStride and Damiao motor controllers also means LeRobot can now talk to higher-torque, professional-grade actuators instead of relying solely on the Dynamixel and Feetech ecosystem it grew up on.

On the model side, six new policies arrived, and the most interesting is Pi0-FAST, which brings autoregressive vision-language-action models back into the mix using a Gemma 300M action expert and frequency-space tokenization. Pair that with Real-Time Chunking, a technique from Physical Intelligence that blends new predictions into in-progress action chunks instead of waiting for a full chunk to finish, and you get policies that react faster in the real world, where latency actually costs you. Wall-X and X-VLA add Qwen2.5-VL and Florence-2 backbones respectively, SARM tackles long multi-stage tasks by modeling stage-aware progress instead of a single linear reward, and new PEFT support means LoRA fine-tuning of big VLAs is now built into the training pipeline rather than something you hack in yourself.

Data collection got the kind of fix that sounds boring until you've actually recorded robot episodes: streaming video encoding kills the dead time between episodes by encoding frames as they're captured, with automatic hardware encoder detection. Hugging Face is also claiming 10x faster image training and 3x faster encoding from fixed data bottlenecks and default parallel encoding. And EnvHub now lets you load simulation environments straight from the Hub the same way you'd load a dataset, while a new NVIDIA IsaacLab-Arena integration brings GPU-accelerated, massively parallel manipulation environments into the training loop.

Underneath all of it, the codebase moved to Python 3.12 and Transformers v5, and third-party policies can now be installed as separate pip packages without touching LeRobot's core, mirroring the plugin system it already built for hardware. It's the kind of unglamorous infrastructure work that doesn't make headlines but determines whether a project like this can keep absorbing 50 new contributors per release without falling apart.