Human-to-Robot Transfer via Bridging Representation and Action Space in VLAs

Human-to-Robot Transfer via Bridging Representation and Action Space in VLAs. By Junhyeong Park and four others, from KAIST and RLWRLD. Published on the third of November.

Pretrained Vision-Language-Action policies, or VLAs, are a strong place to start a deployment. But adapting one to a particular robot in a particular environment still takes embodiment-specific demonstrations, and those are costly to gather. Human demonstrations are far easier to collect: a person can simply show the task. Using them directly is the hard part, because the gap between a human body and a robot's is substantial. It is harder still when the VLA was pretrained mostly on robot data, since then the policy has learned from the one kind of body and is now being shown the other.

H2RB, the Human-To-Robot Bridge, is a co-training framework that adapts a pretrained VLA from human and robot demonstrations that are unpaired. Nothing has to be recorded twice, once by a person and once by the robot. Each kind of demonstration can be gathered on its own and both are trained on together. The framework treats the embodiment gap as showing up at two interfaces, representation and action, and it builds a bridge at each.

For representation, geometry-anchored alignment grounds human and robot observations in a compatible representation, so the two are seen in terms the policy can relate. For action, a decoupled-decoder tokenizer shares action latents between the two domains, without forcing both of them through a single action mapping. That is a balance: the domains meet in a common action space, but neither is made to act through the other's mapping. Together, these let human demonstrations provide useful behavioral supervision while the pretrained VLA keeps its rich visual, semantic and control priors.

The tests are real-world tasks. Object transfer performance rises from eight point zero to fifty-one point eight. Skill transfer scores sixty-two point seven, in settings where the co-training baselines totally fail to transfer skills demonstrated by humans. That second result is the sharper contrast: for those skills, the baselines carry nothing across from the human demonstrations, and H2RB does. The benefit persists under object-placement shifts as well, where out-of-distribution transfer improves from six point zero to forty-five point three.

Robot data is the expensive ingredient here, and human demonstration is the cheap one. H2RB is offered as a way for the cheap one to serve as a scalable supervision source, so that fitting a pretrained robotic policy to a new setting leans less on costly data collected on the target robot itself.