Fac-TDMPC: Learning an Efficient and Robust Factored World Mdoel for Robot Planning

Yuan Zhang, Jianhong Wang, Joschka Boedecker

Keywords: Factored World Model, Decentralized Planning, Model-based Reinforcement Learning

TL;DR: Learning a factored world model on latent space and enabling decentralized planning

Abstract: Model-based reinforcement learning (MBRL) achieves high sample efficiency in robotics, but existing methods such as TDMPC suffer from high planning latency due to the combination of a centralized world model and model predictive control as planning methods, thus limiting real-time deployment in high-dimensional action spaces. We propose Fac-TDMPC, which learns a factored latent-space world model and thus enables decentralized MPC, which largely reduces the optimization complexity in theory. Experiments on high-dimensional robotic control tasks show that Fac-TDMPC achieves substantial planning speedups, while maintaining control performance and improving robustness to action noise.

Submission Number: 11