CuraEval: Evaluating Demonstration Curation on Mixed-Skill Robot Data

CuraEval provides a framework for evaluating demonstration curation in robot imitation learning by comparing selection methods against random subsets and full-data training. Testing on Robomimic data shows that selecting the shortest 50% of demonstrations improves Diffusion Policy success rates from 70.7% to higher benchmarks in square tasks.