Human demonstrations have become a staple for pretraining vision-language-action models, but the new arXiv paper argues they are rarely used for the kind of post-training that makes robots work on a specific body or task. The authors propose a shift in approach: instead of retargeting the human hand to a robot gripper, retarget the object being manipulated. This, they suggest, could make the same demonstration data useful for any robot embodiment.
The paper, titled "Human Demonstrations as Post-Training Data for Any Robot Body: Retargeting the Object, Not the Hand," is aimed at scalable industrial deployment, where task-oriented training on a target platform matters as much as broad pretraining. The abstract is brief, so details on the method or experimental results are not yet available from the source.
If the approach holds up, it could address a bottleneck in robot learning: the gap between abundant human video data and the need for embodiment-specific, task-focused fine-tuning. The core idea is to preserve the object's trajectory and interaction, which transfers more naturally across different robot morphologies than trying to mimic a human hand's motion exactly.