GRASP proposes three innovations making long-horizon planning with world models practical: lifting states for parallel-in-time optimization, stopping brittle state gradients while keeping action gradients, and periodic sync refinement. Achieves leading success rates at long horizons on Push-T.
BLOG