Deep Reinforcement Learning With Action Masking for Differential-Drive Robot Navigation Using Low-Cost Sensors

Driving a wheeled differential-drive robot to a target can be a complicated matter when trying to also avoid obstacles. Usually, such robots employ a variety of sensors, such as LiDAR, depth cameras, and others, that can be quite expensive. To this end, in this paper, we focus on a simple differential-drive wheeled robot that uses only inexpensive ultrasonic distance sensors and touch sensors. We propose a method for training a Reinforcement Learning (RL) agent to perform robot navigation to a target while avoiding obstacles. In order to increase the efficiency of the proposed approach we design appropriate action masks that can significantly increase the learning speed and effectiveness of the learned policy. As we experimentally demonstrated, the proposed agent can robustly navigate to a given target even in unknown procedurally generated environments, or even when denying part of its sensor input. Finally, we show a practical use-case using object detection to dynamically search for, and move to objects within unknown environments. The code used for conducted experiments is available online on Github.