Further, reasoning about key points instead of raw pixels allows Point Policy to generalize to novel object instances and exhibit robustness to heavy scene variations, all while requiring at most 30 demonstrations per task.

Despite having no access to robot demonstrations, Point Policy exhibits an 88% success rate across 8 real-world tasks, a 75% improvement over baselines.

Point Policy uses sparse key points to represent both human demonstrators and robots, bridging the morphology gap. The scene is encoded through semantically meaningful key points from minimal human annotations.