Yann LeCun's retweeted world model planner GRASP code open-sourced

MeNews · 2026-04-05T02:29:55+00:00

On April 1st, the GRASP code was open-sourced. It is a parallel stochastic gradient planner used for learning world models, optimizing actions and intermediate states. GRASP achieves efficient planning through dynamic consistency penalties, features multiple key designs, is suitable for various environments, and can be integrated with other models.

MeNews

2026-04-05 02:29:55

Abstract generation in progress

ME News update: On April 1 (UTC+8), the GRASP world-model planner code released by Michael Psenka and reposted by Yann LeCun has been open-sourced. GRASP is a parallel stochastic gradient planner for learning world models. Its core idea is to simultaneously optimize actions and intermediate “virtual” states in an improved trajectory space. It uses a dynamics-consistency penalty to ensure alignment with the learned dynamics model, and enables efficient parallel planning. Key design elements include introducing virtual states that can be computed in parallel, exploring the state space via noise, using a stop-gradient technique to prevent gradient flow into the state inputs of high-dimensional visual models, and combining goal shaping with periodic synchronized steps of serial unrolling to ensure stability. The planner is built on the DINO-WM project, but can also be integrated with other world models (such as LeWorldModel and JEPA-WMs) via adapters. It is applicable to a variety of environments including PushT and TwoRoom. The codebase provides implementations of the core algorithms and adapter installation scripts. (Source: InFoQ)

View Original

This page may contain third-party content, which is provided for information purposes only (not representations/warranties) and should not be considered as an endorsement of its views by Gate, nor as financial or professional advice. See Disclaimer for details.

1 Likes