I-Robot: Identifying Robotic and Human Motion in Humanoids

Commercial humanoid robots now closely match human body proportions and movement, and ordinary clothing is often enough to make them visually indistinguishable from a person. This visual similarity introduces new risks: a humanoid can be mistaken for a human in surveillance footage, enabling the fabrication of false alibis. However, no existing method is designed to detect such cases. We address this problem by determining whether a motion sequence corresponds to a human or a humanoid, using only temporal pose information without relying on visual appearance cues. We construct HvH (HumanVsHumanoid), a dataset spanning 38 humanoid platforms and 11 shared action classes, and propose I-Robot, a dual-branch model that processes raw pose sequences and their temporal differences in parallel and fuses them via learned per-channel attention. I-Robot consistently outperforms standard sequence and skeleton classifiers on HvH.