Gradient-based Long-term World Model Planning: GRASP Method
This paper introduces GRASP, a new gradient-based planner for solving long-term planning problems. GRASP improves the robustness and efficiency of long-term planning by elevating trajectories to virtual states, adding randomness, and reshaping gradients.
Why it matters: This research proposes GRASP, a new gradient-based planner for solving long-term planning problems, providing new ideas for the application of world models.