đ€ Could watching humans help robots learn how to move?
Researchers have developed EgoExplorer, a wearable system that captures first-person video alongside whole-body movement, detailed finger motion and tactile information.
The resulting dataset contains 100 hours of human activity across 2,715 recordings, covering laboratories, industrial spaces and everyday environments.
Researchers report a 71.9% reduction in body-point error compared with their strongest evaluated baseline, while early experiments transferred human movements to a simulated Unitree G1 humanoid robot.
But there is an important limitation: the robot demonstration was conducted only in simulation, and real-world humanoid learning remains unverified.
The next step could be finding out how far this kind of multimodal human data can take physical robots.
