DexHoldem: Playing Texas Hold'em with Dexterous Embodied System

Current policy models struggle with DexHoldem's complex demands. Even top models achieve only 47.5% scene-preserving success, highlighting the gap between task completion and the precision required to maintain a usable tabletop state for long-horizon tasks.

The benchmark exposes significant challenges in agentic perception. The best perceiver achieves only 34.3% strict full-state accuracy, with routing-critical chip-state fields being particularly unreliable, peaking at 45.8% for current bet chips.

System-level case studies demonstrate how perception and policy errors accumulate in closed-loop deployment. Repeated waiting, recovery dispatches, and human-help requests highlight the compounding reliability gap in long-horizon interactions.

The findings from DexHoldem are directly applicable to developing more robust and reliable AI systems for various enterprise scenarios involving precise manipulation and dynamic environment interaction. The challenges in scene-preserving execution and fine-grained perception are critical hurdles for automation in logistics, manufacturing, and service robotics. Overcoming these will enable a new generation of adaptable and intelligent robotic solutions.

The findings from DexHoldem are directly applicable to developing more robust and reliable AI systems for various enterprise scenarios involving precise manipulation and dynamic environment interaction. The challenges in scene-preserving execution and fine-grained perception are critical hurdles for automation in logistics, manufacturing, and service robotics. Overcoming these will enable a new generation of adaptable and intelligent robotic solutions.