Blog.

The converter CLI is now fully compatible with LeRobot v2.1, including the new episode-metadata schema and the multi-camera observation layout. Migration guide included; existing v1.x datasets are auto-upgraded on load.

We've been using the V0 → V3 validation ladder internally for a while. This post is the first public writeup of the definitions, why each level exists, and how we score existing datasets against it.

Nia Robotics Lab at KTH contributed 340 new edge cases covering low-light warehouse scenarios. The contribution went through our V2 validation protocol; details on the process, review, and merge below.

We've drafted a formal protocol describing how benchmark runs on Smiling Buddha datasets should be reported. This post summarises the protocol and opens a public comment period.

Our first perception benchmark is out of beta. Long-horizon object tracking in cluttered environments, with a reference baseline, protocol, and a public leaderboard.

Short position note on why generic web-scale data pipelines are structurally insufficient for embodied models, and what a physical-AI-native data layer would look like. Companion to the position paper on the Research page.