Research
120d agoGRASP Enables Long-Horizon Planning For World Models
GRASP lifts trajectories into virtual states to parallelize optimization across time. This gradient-based planner solves the instability issues typically found in learned dynamics. By adding stochasticity directly to transitions, it handles complex environments more reliably.
The Signal
Practitioners can now execute longer-horizon planning without the computational collapse common in traditional world model gradients.