HumanPlus: whole-body humanoid robot control from egocentric human video
In one sentence Stanford presents HumanPlus, which maps third-person human demonstrations to whole-body robot actions with 40% success on novel tasks. No teleoperation, no robot-specific data collection — just watching humans.
Teaching a humanoid robot to move like a human has traditionally been a huge problem: you need to collect teleoperation data with expensive equipment, or manually write every single movement. Stanford's HumanPlus finds a completely different path: it simply watches videos of people doing the same things.
The system uses an egocentric camera (as if mounted on the robot's head) to film a person performing a task. It then automatically converts human movements into robot commands, accounting for differences in proportion and kinematics between human and robot bodies. This process is called "retargeting."
With this approach, the robot learns to do complex things like picking up objects, opening drawers, and manipulating tools, achieving 40% success on novel tasks it has never seen before. Not perfect, but extraordinary considering no robot-specific demonstrations were collected.
The real breakthrough is scalability: filming a person doing something requires only a camera and a few minutes. This dramatically lowers the data collection cost and suggests that in the future robots could learn simply by observing humans in everyday life.
Companies
Stanford
Tools
—
Tags
Sources