As I have argued, the winning combination would be planning-based methods using trained world models.
But with such simple kinematics/dynamics and relatively simple grasping tasks, trainable world models are superfluous.

→ View original post on X — @ylecun,