Deep Reinforcement Learning for Mobile Robot Navigation

While navigation is arguable the most important aspect of mobile robotics, complex scenarios with dynamic environments or with teams of cooperative robots are still not satisfactory solved yet. Motivated by the recent successes in the reinforcement learning domain, the application of deep reinforcement learning to robot navigation was examined in this paper. In particular this required the development of a training procedure, a set of actions available to the robot, a suitable state representation and a reward function. The setup was evaluated using a simulated real-time environment. A reference setup, different goal-oriented exploration strategies and two different robot kinematics (holonomic, differential) were compared in the evaluation. In a challenging scenario with obstacles at changing locations in the environment the robot was able to reach the desired goal in 93% of the episodes.