Event

H2RBench Human-to-Robot Transfer Benchmark Released

1 reports · 1 sources

Chuyang Xiao and other authors propose and release H2RBench, a Real2Sim benchmark for evaluating "human-to-robot" (H2R) transfer methods. The benchmark establishes a standardized protocol based on real human video demonstrations and simulated robot demonstrations, covering four categories of manipulation tasks, and is used to compare the ability of different methods to leverage human demonstration data. The authors report that testing with H2RBench found clear differences in how various methods scale as the number of human demonstrations increases; performance in simulation generally predicts real robot performance, with a Pearson correlation coefficient of r=0.89, a Spearman correlation coefficient of ρ=0.85, and a mean maximum rank violation (MMRV) of 0.06 across method-task configurations. These correlations are evaluation results the authors present based on this benchmark.