UBTech Tests GPT-6 Astra on Robots: Direct Grasping Failed, Stacking Succeeded

The split result is the real signal: Astra's direct grasping failed on hardware while stacking succeeded only through UBTech's own Thinker-VLA and Thinker-WM stack, suggesting general-purpose models still need vendor-specific scaffolding to close the loop on physical tasks. That distinction matters for anyone evaluating embodied-AI claims, since headline model branding can obscure which component actually drives performance. The open question is whether these are one-off test outcomes or a pattern that shapes how robot makers position foundation models versus in-house control layers.