OpenAI researcher probes why general multimodal model Astra beats specialized robotics models

willdepue · x · 2026-11-01

OpenAI researcher Will Depue revised his quick take on Astra's strong robot task performance, suggesting "computer use" transfer from video input may explain its rapid object control. He then asks why specialized robotics models underperformed: compute allocation can't explain it (robotics models likely spent equal or more flops on robotics-like/multimodal data), Astra's multimodal encoders are unlikely superior, and chain-of-thought may help robotics tasks.

Related event: OpenAI's Astra Beats Specialized Robot Models Zero-Shot, Sparking Transfer Debate(12 posts)→

Original post →

More from Embodied

Embodied channel →