A single world model serving prediction and control across bodies could simplify robot learning.
GreenWAM is a world foundation model that both predicts how a robot's actions change a scene and chooses those actions, aiming to work across different robot bodies.