Human Video Datasets Enhance Robot Learning Efficiency
Research explores the use of robotized human videos for training VLA policies.
Recent research investigates the potential of robotized human video datasets as a cost-effective alternative to traditional real-robot data. The study introduces a robotization pipeline that transforms diverse human videos into robot-aligned observations and action trajectories, leading to the creation of the HuRo dataset, which includes approximately 630,000 robotized episodes.
Results indicate that scaling up robotized pretraining significantly boosts task completion rates across various manipulation tasks, with improvements noted from 51.5% to 80.3%. The findings also highlight the advantages of visual robotization in enhancing out-of-distribution robustness.