Transfer learning for reinforcement learning on a physical robot

摘要

As robots become more widely available, many capabilities that were once only practical to develop and test in simulation are becoming feasible on real, physically grounded, robots. This newfound feasibility is important because simulators rarely represent the world with sufficient fidelity that developed behaviors will work as desired in the real world. However, development and testing on robots remains difficult and time consuming, so it is desirable to minimize the number of trials needed when developing robot behaviors. This paper focuses on reinforcement learning (RL) on physically grounded robots. A few noteworthy exceptions notwithstanding, RL has typically been done purely in simulation, or, at best, initially in simulation with the eventual learned behaviors run on a real robot. However, some recent RL methods exhibit sufficiently low sample complexity to enable learning entirely on robots. One such method is transfer learning for RL. The main contribution of this paper is the first empirical demonstration that transfer learning can significantly speed up and even improve asymptotic performance of RL done entirely on a physical robot. In addition, we show that transferring information learned in simulation can bolster additional learning on the robot.