Human fingers and robotic fingers have different geometry, so copying a hand trajectory^[1] does not guarantee the same contact with an object. Haozhe Zhang and collaborators propose TouchBridge, a combined hardware-and-software system with a wearable human demonstration interface and a corresponding robotic interface. Both capture visual and fingertip tactile information to establish correspondence across different hand designs.

Its Tac-Retarget component aligns fingertip contact patches and their spatial relationships using measured touch. The authors report that this conversion does not require an object mesh, an object-pose estimate or a physics simulator. With visual synthesis and camera-anchored action conversion, the transformed examples are used to fine-tune a pretrained model that maps visual observations and instructions to robot actions.

Across four physical contact-rich tasks, policies trained solely on converted human demonstrations reached 72.5% average success. An equal number of robot demonstrations collected through direct teleoperation yielded 77.5%, while human data collection was seven times faster. These results are promising for collecting examples efficiently, but cover specific tasks, hands and training conditions, not every object or robotic platform.

The most revealing question is the difference between repeating a learned example and behaving well in a genuinely new situation. Generalization^[2] means coping with conditions that were not identical to the training examples. One striking video cannot tell us how many attempts failed. It helps to ask what the held-out tests looked like, whether objects and lighting changed, who selected the demonstrations and which baselines were compared. Those questions turn a fascinating result into a testable claim rather than a promise of a robot that can already do everything. Sevenfold faster collection does not remove the need to test contact on target objects. The 72.5% and 77.5% results matter alongside failure costs.