Simulation and Transfer of Reinforcement Learning Algorithms for Autonomous Obstacle Avoidance

The explicit programming of obstacle avoidance by an autonomous robot can be a computationally expensive undertaking. The application of reinforcement learning algorithms promises a reduction of programming effort. However, these algorithms build on iterative training processes and therefore are time-consuming. In order to overcome this drawback we propose to displace the training process to abstract simulation scenarios. In this study we trained four different reinforcement algorithms (Q-Learning, Deep-Q-Learning, Deep Deterministic Policy Gradient and A synchronous Advantage-Actor-Critic) in different abstract simulation scenarios and transferred the learning results to an autonomous robot. Except for the Asynchronous Advantage-Actor-Critic we achieved good obstacle avoidance during the simulation. Without further real-world training the policies learned by Q-Learning and Deep-Q-Learning achieved immediately obstacle avoidance when transferred to an autonomous robot.