Research2h ago
GRASP Enables Long-Horizon World Model Planning
GRASP optimizes trajectories by lifting them into virtual states to parallelize computation across time. This gradient-based planner solves the instability issues common in long-horizon dynamics. Researchers at BAIR added stochasticity directly to the planning process to improve robustness. Practitioners can now execute complex, multi-step tasks without the typical gradient explosion in world models.