H2RB: Bridging Human-Robot Embodiment Gaps in VLA Training
November 2, 2026
H2RB is a co-training framework designed to adapt pretrained Vision-Language-Action (VLA) policies from unpaired human and robot demonstrations. It employs geometry-anchored alignment and a decoupled-decoder tokenizer to bridge the gap between human observations and robot action spaces.
HOW THIS AFFECTS YOU
●
builderYou can leverage human demonstrations to adapt VLA policies to new robots more cheaply.
●
researcherThe decoupled-decoder tokenizer offers a new way to share action latents across different embodiments.