A way for robots to learn doable actions from videos.
It is like a robot apprentice binge-watching DIY clips. It sees someone tighten a screw, then goes to practice.
It cuts the cost of collecting teaching demos. It can help home robots, factory robots, and robot hands.
Imitation Learning
It turns public videos into demos for Imitation Learning.
Embodied Data
Video is cheap Embodied Data, but it often lacks action labels.
VLA
The learned link between seeing and acting is often used to train VLA.
Embodied AI
It lets Embodied AI learn from videos of human activity.