What's this about?
PointWAM learns how hands and scenes co-evolve as 3D point trajectories, then translates forecast human hand motion into robot control.
What exactly is a world-action model?
It jointly predicts what the environment will do and which action the robot should take, so anticipated physical consequences guide control.
Why does this matter?
Explicit 3D contact geometry may transfer human demonstrations better than flat video pixels. The reported gains are large, but depend on the chosen robot tasks and retargeting system.