Forward Model Learning for Motion Control Tasks

In this work, we study the capabilities and limitations of forward model learning agents and their applications to motion-control tasks. Forward model learning agents learn to approximate the environment dynamics to apply planning algorithms for action-selection. While previous work has shown that forward model learning agents can efficiently learn to play simple video games, we extend their applicability to domains with continuous state and action spaces. Our experiments show that such agents are quickly able to learn an approximate model of their environment, which suffices to solve several simple motion-control tasks. Comparisons with deep reinforcement learning further highlight the sample efficiency of forward model learning agents.