GRASP: New Parallel Planner for World Models скачать в хорошем качестве

GRASP: New Parallel Planner for World Models

In this AI Research Roundup episode, Alex discusses the paper: 'Parallel Stochastic Gradient-Based Planning for World Models' Researchers introduce GRASP, a robust and highly parallelizable planner designed for vision-based world models. By treating states as optimization variables with soft dynamics constraints, the method enables efficient parallel computation for long-horizon control tasks. To improve exploration and avoid local optima, the planner incorporates stochasticity into the state representation. The system modifies gradient structures to handle high-dimensional visual inputs while only requiring action-input gradients. Experiments show that GRASP significantly outperforms existing planning algorithms in complex video-based environments. Paper URL: https://arxiv.org/pdf/2602.00475 #AI #MachineLearning #DeepLearning #WorldModels #Robotics #ComputerVision #MetaFAIR