LeRobot v0.5.0リリース、ロボット学習基盤を全方位で拡張 LeRobot v0.5.0: Scaling Every Dimension

Hugging Face has shipped LeRobot v0.5.0, a major update to its open-source robotics learning library that aims to be a common substrate for imitation learning, reinforcement learning, and real-world robot deployment. The release touches nearly every layer of the stack, from data handling to policies to hardware integration.

The headline change is a unified processor pipeline that standardises how observations and actions are transformed between datasets, training, and inference. In practice, mismatches between the preprocessing used at train time and at deploy time are a frequent source of silent failures in robot learning, so making this pipeline explicit and reusable should reduce a class of subtle bugs. Alongside this, async inference has been added, decoupling policy computation from the control loop so that high-frequency robots are not bottlenecked by model latency.

On the policy side, v0.5.0 expands support for state-of-the-art approaches including SmolVLA, a compact vision-language-action model developed within the LeRobot ecosystem, as well as ACT, Diffusion Policy and Pi0-style architectures. A reinforcement learning workflow has also been integrated, reflecting a broader research trend of using RL to fine-tune policies first bootstrapped via imitation learning. The LeRobotDataset format, tightly coupled with the Hugging Face Hub, has been refined to scale better to larger episodic datasets.

Hardware coverage continues to grow, spanning low-cost educational arms such as the SO-100 and SO-101 up to more capable bimanual setups. This dovetails with Hugging Face's acquisition of Pollen Robotics and ongoing work around the Reachy 2 humanoid, suggesting a deliberate strategy of co-developing software and accessible hardware rather than treating them as separate concerns.

The wider context matters here. The last year has seen a surge of large vision-language-action models, including Physical Intelligence's Pi0, Google DeepMind's RT-2 and Gemini Robotics work, and NVIDIA's GR00T initiative for humanoids. These systems are powerful but often hard to reproduce or compare, and evaluation in robotics remains notoriously inconsistent. LeRobot is positioning itself as something like a transformers library for robotics: a place where heterogeneous policies, datasets and embodiments can be expressed through shared abstractions and benchmarked on common ground.

Whether the project can sustain that ambition will depend on community uptake and on how cleanly the new processor and async-inference abstractions hold up as more exotic robots and policies are added. For now, v0.5.0 looks like a substantive step in that direction, and it is likely to lower the barrier for researchers and hobbyists who want to move from notebook-scale experiments to actual hardware without rewriting their stack from scratch.