DexHoldem: Playing Texas Hold'em with Dexterous Embodied System

Researchers introduced DexHoldem, a real-world Texas Hold'em benchmark for dexterous embodied agents.

The dataset includes 1,470 teleoperated demonstrations spanning 14 manipulation primitives.

It evaluates two tracks: physical primitive execution and structured game-state perception.

Results show large model differences in task completion, scene preservation, and perception accuracy.

Case studies show how small errors can cascade in full closed-loop agent deployment.