PRISM expands four real videos into 256 counterfactual variants for humanoid robot training

A contributor to PRISM says its real-to-sim-to-real approach expands four real videos into 256 counterfactual variants to train humanoid robots to move and handle objects. They say it recovers robot-object trajectories and trains one policy that generalizes across objects, scales and spatial layouts. The work was done at Amazon FAR.