DexHoldem: Playing Texas Hold'em with Dexterous Embodied System

Real-world embodied systems must perceive a changing scene, select a legal action, execute contact-rich manipulation, and verify that the scene remains usable. We introduce **DexHoldem**, a real-world Texas Hold'em benchmark that evaluates these capabilities in one ShadowHand tabletop environment. The benchmark provides 1,470 teleoperated demonstrations across 14 card and chip primitives, a shared multi-view policy interface, an agentic-perception benchmark for structured game-state recovery, and a closed-loop evaluation protocol. Across 80 physical primitive trials, $\pi_{0.5}$ reaches 61.2% task completion and $\pi_{0.5}$/$\pi_0$ reach 47.5% scene-preserving success. On 36 perception problems, the best strict full-state accuracy is 34.3%, while the best field-wise average is 66.8%; chip-state fields remain the main bottleneck. Three hand-level case studies show how waiting, recovery, and verification costs accumulate when perception and dexterous execution are composed. DexHoldem provides a compact physical testbed for interactive, multimodal embodied deployment.