Human-to-Robot Transfer via Bridging Representation and Action Space in VLAs
H2RB raised reported real-world object-transfer performance from 8.0 to 51.8 using unpaired human and robot demonstrations.
The framework adapts pretrained vision-language-action policies by aligning human and robot observations while sharing action representations through separate decoders. It achieved a skill-transfer score of 62.7 where co-training baselines failed to transfer human-demonstrated skills. Under shifts in object placement, transfer performance rose from 6.0 to 45.3. The authors present the approach as a way to reduce reliance on costly demonstrations collected on the target robot.
HF Daily Papers' note
The framework adapts pretrained vision-language-action policies by aligning human and robot observations while sharing action representations through separate decoders. It achieved a skill-transfer score of 62.7 where co-training baselines failed to transfer human-demonstrated skills. Under shifts in object placement, transfer performance rose from 6.0 to 45.3. The authors present the approach as a way to reduce reliance on costly demonstrations collected on the target robot.
HF Daily Papers' note
score 5