Every model.
Every answer.
Compare recorded predictions against human labels, inspect disagreements, and browse the sampled LeRobot catalogue.
Compare the models
862 episodes have a clear outcome. The 65 ambiguous community clips remain available for inspection and are excluded from scored metrics.
Inputs and comparison methods
Released RoboJev reads ten evenly spaced front-camera frames plus final views from other cameras when available. Frontier LLMs answered zero-shot from four front-camera frames. clip-v1 means four frames, dense-v1 means ten plus other views, and video means whole-video input.
SO-101: 268 held-out runs, with RoboJev trained on other ArmnetBench sessions. Unseen labs: 459 runs from eight labs outside training. Community: 200 clips labelled by one person, 135 with clear outcomes scored. Missing results are shown as a dash.
A sample of robot evaluations
68,555 sampled episodes, totalling about 501 hours. This is a catalogue, rather than a complete census of Hugging Face. Earlier four-frame predictions are omitted; the current release has not regraded the whole catalogue.
Download catalogue JSONLLoading recorded results…