PRISM boosts robot planning by 35% with lightweight action priors

In robotics, learned world models help agents simulate future states, but planning requires efficiently generating candidate actions. Existing approaches often rely on arbitrary search or expert demonstrations, discarding state-conditioned confidence, or use heavyweight external encoders like large VLMs. PRISM, from a team of researchers, argues this complexity is unnecessary. Building on a standard JEPA-style latent world model, they attach a lightweight MLP directly to its frozen encoder to predict a Gaussian prior. At plan time, this prior is fused into the planner's sampling distribution via a precision-weighted Product-of-Gaussians update — a closed-form, parameter-free integration that steers sampling where confident and yields control elsewhere.

PRISM achieves 35 percentage point improvement on the Cube task and 32 on PushT over vanilla model-based MPC, with no meaningful inference overhead added. The method is task-agnostic, requires only the same data used to train the world model, and avoids architectural bloat from separate visual encoders or large vision-language models. This makes PRISM a practical, efficient solution for continuous control in robotics, enabling better planning without additional model complexity or computational cost.