DexHoldem: Playing Texas Hold'em with Dexterous Embodied System

Overview

DexHoldem is a real-world embodied-system benchmark that brings dexterous manipulation to the Texas Hold’em table. It evaluates how a robotic system combines tabletop execution, agentic perception, and decision routing on a ShadowHand/UR10e platform.

What we explore

• Dexterous execution: card and chip manipulation, supported by 1,470 teleoperated demonstrations across 14 primitives.
• Scene understanding: multi-view observations and 36 structured perception problems.
• Closed-loop behavior: case studies connecting scene parsing, action routing, and dexterous policy execution.

My contribution

I co-proposed and led the project, built the data-collection infrastructure, maintained the hardware platform, and designed policy, perception, and system-level evaluations. I also benchmarked pretrained and from-scratch policy families, examining task completion, scene preservation, data scaling, and real-world failure modes.