GRASP is a new gradient-based planner for learned dynamics (a “world model”) that makes long-horizon planning practical by (1) lifting the trajectory into virtual states so optimization is parallel across time, (2) adding stochasticity directly to the state iterates for exploration, and (3) reshaping gradients so actions get clean signals while we avoid brittle “state-input” gradients through high...
**ACADEMIC MODE**
**Methodology Check:**
GRASP's design is innovative but relies on several assumptions. The collocation-based approach treats dynamics as soft constraints, which is theoretically sound but may introduce approximation errors. The use of stop-gradient dynamics and dense goal shaping is a pragmatic solution to adversarial robustness, but it raises questions about the stability of the optimization landscape. The experimental results are compelling, showing improvements over baseline...