Reinforcement learning for robots through efficient simulator sampling

[摘要] Reinforcement learning (RL) has great potential in robotic systems as a tool for developing policies and controllers in novel situations. However, the cost of realworld samples remains prohibitive as most RL algorithms require a large number of samples before learning near-optimal or even useful policies. Simulators are one way to decrease the number of required real-world samples, but imperfect models make deciding when and how to trust samples from a simulator difficult. Two frameworks are presented for efficient RL through the use of simulators. The first framework considers scenarios where multiple simulators of a target task are available, each with varying levels of fidelity. It is designed to limit the number of samples used in each successively higher-fidelity/cost simulator by allowing a learning agent to choose to run trajectories at the lowest level simulator that will still provide it with useful information. Theoretical proofs of this framework;;s sample complexity are given and empirical results are demonstrated on a robotic car with multiple simulators. The second framework focuses on problems represented with continuous states and actions, as are common in many robotics domains. Using probabilistic model-based policy search algorithms and principles of optimal control, this second framework uses data from simulators as prior information for the real-world learning. The framework is tested on a propeller-driven inverted pendulum and on a drifting robotic car. These novel frameworks enable RL algorithms to find near-optimal policies in physical robot domains with fewer expensive real-world samples than previous transfer approaches or learning without simulators.

[发布机构] Massachusetts Institute of Technology