98,000 simulated evaluations reveal robots making surprisingly clever physical reasoning

tomssilver · x · 2026-09-25

Tom Silver (Princeton) introduces AgenticGenTAMP: with collaborators he ran 98,000 sandboxed evaluations of Astra robots across 28 simulated environments, with a new preprint released.

His key observation: unexpected robot behavior usually means a bug, but this project is one of the few times in his career he's seen robots act "unexpectedly clever" in ways nobody on the team anticipated. The thread walks through examples—expected vs. actual behavior, favorite 2D cases—showing emergent physical reasoning from coding-agent-synthesized planning policies.

Related event: 98,000 Simulated Trials Show Astra Robot's Surprising Physical Reasoning(3 posts)→

Original post →

More from Embodied

Embodied channel →