Visuomotor FusionIntermediate
Register-Routed Delayed Fusion: Rewiring Shortcut-Prone Observation Fusion in Visuomotor Imitation
This arXiv preprint introduces Register-Routed Delayed Fusion (RRDF), a Transformer observation-fusion schedule for visuomotor imitation that blocks direct proprioception-to-vision attention and routes cross-modal information through learned registers. On three real-robot tasks, RRDF improves nominal success rates from 55%, 65%, and 50% for dense ACT to 100%, 100%, and 85%; under held-out object positions, it improves from 30%, 30%, and 0% to 95%, 100%, and 20%. The policy also shows lower state-to-image action sensitivity, suggesting more robust visual responsiveness in imitation-learned manipulation.