
Scaling Bimanual Household Manipulation from 1,500 hours of Demonstrations to On-Policy Corrections
The authors release 1,500 hours of bimanual household manipulation data and train XR-2, a 5B-parameter VLA. On clothes folding, expert data lifts success from 34% to 84% before saturating, and three rounds of on-policy DAgger corrections then raise it from 58% to 93%.
Jiafeng Xu, Qi Li, Yan ShenSep 3, 2026
Bimanual ManipulationVision-language-action (VLA)Imitation LearningSep 3, 2026