Research
118d agoGRASP Enables Long-Horizon World Model Planning
The GRASP planner optimizes trajectories by lifting them into virtual states, allowing parallelization across time. This approach solves the instability common in gradient-based planning for learned dynamics. By adding stochasticity directly to the trajectory, it avoids local minima.
The Signal
Researchers can now execute complex, long-term plans without the typical computational collapse of world models.