DexHoldem: Playing Texas Hold'em with Dexterous Embodied System

Researchers introduced DexHoldem, a real-world benchmark for dexterous embodied robotics built around poker-table manipulation on a ShadowHand system.

The dataset includes 1,470 teleoperated demonstrations across 14 manipulation primitives, along with physical policy and agentic perception evaluations.

In the reported results, pi_0.5 achieved the strongest primitive completion, while Opus 4.7 and GPT 5.5 led different perception metrics.

Case studies showed compounding errors in closed-loop deployment, underscoring the gap between isolated skills and reliable end-to-end task performance.