Many robots that learn from human demonstrations move much slower than people, which limits their use in everyday tasks like cooking or cleaning. This happens because current methods for recording demonstrations are slow and awkward, and the learned policies often produce jerky motions that fail at high speed.
The researchers built a robot platform with arms that humans can push directly to demonstrate tasks. The arms are lightweight and compensate for gravity, so the teacher can move them naturally at near-human speed. A learning pipeline then trains the robot from these demonstrations, using a smoothing technique at test time to produce fluid motions and masking to ignore the teacher’s hands in the camera view.
With just 50 examples per task, the robot autonomously performs bimanual jobs, like tossing and catching a ball, folding shorts, or inserting a battery, with over 80% success. Execution time is cut by more than half compared to previous systems, bringing robots closer to practical, human-speed operation.
