Gradient-based Planning for World Models at Longer Horizons

GRASP is a new gradient-based planner for learned world models that enables practical long-horizon planning. It lifts trajectories into virtual states for parallel optimization, adds stochasticity for exploration, and reshapes gradients to avoid brittle signals through high-dimensional vision models. The work was done with Mike Rabbat, Aditi Krishnapriyan, Yann LeCun, and Amir Bar.