BEHAVIOR Challenge Updates

drfeifei · x · 2026-07-14

The 2026 BEHAVIOR Challenge has updated its evaluation setup: this year features only one official track, evaluating models solely using robot embodiment observations, including RGB, Depth, and proprioception.

The reply also details data and baselines: 20,000 human teleoperation demonstrations have been added, totaling 1950 hours—double the scale of 2025. It also provides RGBD observations, robot embodiment sensing, skill/sub-task annotations, and supports strong baselines like pi0.5 and GR00T N1.7.

Related event: Stanford Launches 2026 BEHAVIOR Challenge and BEHAVIOR-1K Benchmark(9 posts)→

Original post →

More from Embodied

Embodied channel →