Event
ACT-Flow proposes an online learning framework for visual insertion
1 reports · 1 sources
Yiqing Yuan et al. propose the ACT-Flow online learning framework, which combines an adaptive Action Chunking Transformer (ACT) with a conditional residual flow policy for correction learning in visual insertion tasks: ACT absorbs successfully executed actions through behavior cloning, and the residual target is recomputed relative to the updated prior. According to the study, candidate scoring is used only during the training phase; at deployment it relies solely on RGB observations and samples one residual per action chunk. On five tasks, the average success rate at 20,000 interaction steps after initialization is 66%, which is 18 percentage points higher than the strongest baseline at that checkpoint; when the residual is disabled at the same checkpoint, ACT's average success rate rises from 16% to 76%.
Report timeline
- ACT-Flow: Learning Visual Insertion Error Correction with Critic-Weighted Residual Flow
ACT-Flow is an online learning framework that combines an adaptive Action Chunking Transformer (ACT) with a conditional residual flow policy, allowing ACT to absorb successfully executed actions through behavioral cloning, with residual targets recomputed relative to the updated prior.
arXiv cs.RO ·