Teaching a robot hand to play poker reveals real challenges

Playing poker requires a robot to do more than execute isolated tricks: it must see cards, choose moves, execute them with a dexterous hand, and leave the table usable. DexHoldem is a real-world benchmark built around Texas Hold'em with a ShadowHand robot, including 1,470 teleoperated demonstrations and tests for both primitive execution and agentic perception. Top policies achieve 61% task completion but only 47.5% on the harder metric of preserving the scene, while leading vision models reach just 34% accuracy on game-state recovery—exposing a gap between visual skills and embodied decision-making.