Research5h ago
GRASP Enables Long-Horizon Planning For World Models
The GRASP planner optimizes trajectories by lifting them into virtual states to enable parallelization across time. This approach solves the computational bottlenecks typical of gradient-based planning in learned dynamics. By adding stochasticity directly to the process, researchers at BAIR improved stability. Practitioners can now execute complex, long-term tasks without exponential cost increases.