People often learn a new task by watching someone demonstrate it. Researchers are exploring whether robots can use a similar source of information. USC recently described Psi 0, a foundation model for humanoid robot learning that combines more than 800 hours of human video with a smaller collection of robot-specific data.

The attraction of human video is scale. There are far more recordings of people reaching, lifting, opening, and placing objects than there are carefully labeled robot demonstrations. But a video of a person is not a direct instruction for a machine. Humans have different joints, balance, grip strength, and viewpoints. The robot must learn the structure of the action while adapting it to its own body.

The reported goal is longer loco-manipulation tasks, meaning the robot can move through an environment while using its arms and hands. That is harder than repeating one isolated motion. The robot has to maintain a plan across several steps and recover if an object is not exactly where the model expected it to be.

This topic creates a useful school project about imitation and adaptation. Students can record a short demonstration of moving an object through a marked path, then describe which parts are universal and which depend on the person performing them. A simple robot can follow the path with fixed commands, while the class discusses what extra sensors would be needed for the robot to respond to variation.

The details are drawn from USC's recent report on humanoid robot learning. The broader lesson for robotics students is that data is not magic; it becomes useful only when engineers decide how to represent, test, and adapt it.

‍