Researchers from Meta and UC Berkeley propose GRASP, a gradient-based planner that stabilizes optimization over large learned world models for long-horizon control.
A research team including Yann LeCun and collaborators from Meta AI and UC Berkeley introduced GRASP, a gradient-based planning algorithm designed to work with large learned world models. The method addresses three core failure modes of long-horizon planning: ill-conditioned optimization, bad local minima from non-greedy structure, and brittle gradients through high-dimensional visual latent spaces. GRASP achieves this by parallelizing trajectory optimization across time via virtual states, injecting stochasticity into state iterates, and reshaping gradients to isolate clean action signals. The work is presented as a research paper with an accompanying project website, not yet deployed in a production system.
If you're building on top of learned dynamics models — for robotics, simulation, or model-based RL — GRASP directly attacks the optimization pathologies that make long-horizon planning unreliable in practice. The core insight is architectural: by lifting trajectories into virtual states and reshaping gradients, you decouple action optimization from the brittle backprop-through-vision problem. This is a practical contribution, not just theory — the failure modes it addresses (vanishing/exploding gradients through high-dim latents, local minima from greedy rollouts) are ones you've hit if you've tried planning over image-based world models.
If you have a world model-based pipeline (e.g. DreamerV3, TDMPC2, or a custom dynamics model), read the GRASP paper's gradient reshaping section and benchmark your current planner's action gradient norms over 20+ step horizons against GRASP's reported values.
Go to claude.ai and start a new conversation
Tags