Research
119d agoGRASP Enables Long-Horizon World Model Planning
The GRASP planner optimizes long-horizon trajectories by lifting them into virtual states for parallel processing. This approach adds stochasticity directly to the optimization process to avoid local minima. Researchers at BAIR developed the method to make gradient-based planning practical for learned dynamics.
The Signal
Practitioners can now execute more complex, multi-step tasks in simulated environments.