Robots can now learn how to handle objects by watching humans do it — no robot training data required.
Researchers introduced Human2Any, a framework that extracts reusable manipulation knowledge from human videos and applies it to robots. Instead of modeling how a human hand moves, it focuses on how objects interact with each other — the motion and contact that actually define a task. That abstraction lets the same learned knowledge transfer across different robot bodies, table layouts, and scenarios. The team validated the approach on a Franka tabletop arm and an RBY-1 humanoid mobile robot.
The core problem Human2Any is attacking is real: human video is cheap and plentiful, but robot demonstration data is expensive and brittle — collected in one setup, useless the moment the scene changes. By anchoring learning to object-object interaction rather than limb kinematics, the framework sidesteps the embodiment mismatch that has stalled this field for years.
Robotics researchers have been chasing this goal since at least the imitation learning wave of the early 2020s, and the graveyard of "transfer learning" papers that never left the lab is long. Human2Any's object-centric framing is a cleaner angle than most — but real-world clutter, lighting variation, and novel object categories will be the test, and a two-robot validation is a start, not a verdict.