A benchmark suite for teaching robot hands complex skills

Dexterous robotic hands can perform manipulation tasks beyond the reach of parallel grippers, but lack standardized benchmarks to measure progress. DexJoCo introduces 11 functionally grounded tasks covering tool use, bimanual coordination, long-horizon execution, and reasoning, with a low-cost data collection pipeline and 1.1K demonstrations. The benchmark includes domain randomization protocols and evaluates modern learning methods across visual randomization, dynamics variation, multi-task training, and action-head adaptation. Empirical analysis reveals common failure modes and robustness limitations in current policies, providing a foundation for advancing dexterous manipulation research. Code and benchmark available online.