Intel / article

Gradient-based Planning for World Models at Longer Horizons

Source: bair.berkeley.edu
Date: Published Apr 20, 2026

Context

A model that predicts future events still needs a planner to choose actions. GRASP tackles this by revising proposed intermediate states and actions together, allowing work on different times in a plan to run in parallel. The Berkeley AI Research explanation describes why this can help when a goal requires a long sequence of moves.

The difficulty is that optimization can exploit errors in a learned simulator. GRASP limits a sensitive feedback path through the simulated state, introduces random changes to explore alternatives, and periodically refines complete simulated trajectories. In the reported Push-T task, it achieved higher success rates than the compared planners at horizons of 50 to 80 steps, while remaining far from solving every attempt.

Key takeaways · AI-distilled

Long-horizon planning through a world model suffers exploding/vanishing gradients because the Jacobian with respect to early actions is a product of per-step Jacobians whose conditioning scales exponentially with the planning horizon.
GRASP's key trick is stopping gradients into the state input of the world model, trusting only action gradients, because states sit on a low-dimensional manifold with adversarial-example-like brittleness in directions orthogonal to it, which the action space doesn't share.
Because the naive stop-gradient objective has a trivial minimum (stay put, then jump to the goal on the last step), GRASP adds a dense per-timestep goal term to every predicted state, balanced against the stop-gradient dynamics loss.
On the Push-T at horizon 80, GRASP hit a 10.4% success rate in a median 58.9s, versus CEM's 2.8%/132.2s and LatCo's 0.0%; at horizon 40, GRASP matched CEM's roughly 60% success rate in about a quarter of the time (8.5s vs 35.3s).
GRASP periodically breaks from its fast lifted-state optimization to run a short refinement phase using the true serial rollout gradient, to keep the lifted trajectory in dynamics actually reachable by a real rollout.

Terms in this piece · Glossary

benchmark — A standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.
grounding — Tying a model's answers to checkable sources — retrieved documents, live data, tool results — instead of letting it answer from memory alone.

Read the source bair.berkeley.edu