Human-to-Robot Transfer via Bridging Representation and Action Space in VLAs
| Source: arXiv AI
Tags: H2RB, Vision-Language-Action Models, Human-to-Robot Transfer, Robot Learning, Imitation Learning
The H2RB framework adapts pretrained vision-language-action robot policies using unpaired human demonstrations, raising real-world object transfer scores from 8.0 to 51.8.
Details
Adapting a pretrained VLA policy to a new robot and environment usually needs costly robot-specific demonstrations. Human demonstrations are cheaper but differ sharply from robot data. H2RB (Human-To-Robot Bridge) co-trains on unpaired human and robot demonstrations. Geometry-anchored alignment grounds both in a compatible representation, while a decoupled-decoder tokenizer shares action latents without forcing both domains through one action mapping. The design lets human demonstrations supply behavioral supervision while keeping the VLA's pretrained priors. Across real-world tasks, H2RB improved object transfer performance from 8.0 to 51.8 and reached a skill transfer score of 62.7.