Google DeepMind Introduces Diffusion Model Predictive Control (D-MPC): Combining Multi-Step Action Proposals and Dynamics Models Using Diffusion Models for Online MPC

We are a community of AI/ ML/Generative AI enthusiasts/researchers/journalists/writers who share interesting news and articles about the applications of AI. 

Machine Learning News

Google DeepMind has introduced Diffusion Model Predictive Control (D-MPC), leveraging diffusion models to combine multi-step action proposals with dynamics models for online MPC. D-MPC has shown superior performance on the D4RL benchmark, adapting well to new dynamics and rewards at runtime. Key components include a multi-step dynamics model, action proposals, and an SSR planner that offers simplified and effective planning. Future work aims to improve sampling speeds and extend capabilities to pixel-based observations.