Robots acting 'unexpectedly clever': AgenticGenTAMP thread showcases emergent behaviors
tomssilver · x · 2026-09-25
In a thread, Tom Silver explains that unexpected robot behavior usually means a bug—but AgenticGenTAMP produced physical reasoning so clever that none of the project members anticipated it, a rare experience in his career. He shares several examples: expected-vs-actual comparisons, his favorite 2D case, and a slower-but-perhaps-more-robust solution. The project accompanies a large-scale sandboxed evaluation preprint: 98,000 evaluations across 28 simulated environments.
Related event: 98,000 Simulated Trials Show Astra Robot's Surprising Physical Reasoning(3 posts)→
More from Fun
- Muse agent files $171 in decade-old refunds, cancels subscriptions, and haggles a TV deal in two days — armand_ruiz · 2026-09-25
- Ex-OpenAI's Miles Brundage: AI industry golden handcuffs are invisible gold that auto-tightens — Miles_Brundage · 2026-09-25
- Solo dev used AI to recreate North America's 1830-1900 railroad boom in a game — randal_olson · 2026-09-25
- Asked Claude to Draw Robots, It Drew a Lion King-Style Scene — _Stocko_ · 2026-09-25
- Opus 5.5 reverse-engineers a battery CAD model from photos in 100 minutes, two prompts — _Stocko_ · 2026-09-25
- AI Persona Refuses to Reveal Its Camera and Keeps 'Acting' Past the Video's End — HarperSCarroll · 2026-09-25