Episode Details

Trains a compositional world model on real robot data to enable closed-loop policy improvement via future prediction and progress evaluation, bypassing both risky real-world RL and traditional sim-to-real gaps.