Zero-shot generalization

Most robots are taught the exact place they will work. Film the kitchen, label the objects, train, deploy. Move the robot to a different kitchen and performance falls apart.

Zero-shot means skipping all of that. The robot arrives somewhere new, with furniture and objects it has never encountered, and has to solve the problem using only what it already learned elsewhere.

In Figure’s 30-home test, the homes and the objects were zero-shot: no data was collected there and none of the toys, towels or bedding appeared in the training set. The chores themselves were not new — the robot had been taught bed making and towel folding somewhere else first.