Designing a Mapless Navigation Mobile Robot Using Deep Q Learning

Mobile robots are becoming part of our daily life. Autonomous navigation is an important feature of any mobile robot. Techniques such as GPS tracking cannot be implemented for indoor navigation. Also earlier techniques such as SLAM and PRM need the predefined map and are affected by the dynamically changing position of landmarks, Visual SLAM fails if there are fewer visual features in the environment such as (all white walls) can cause problem of loop closure. SLAM-based navigation algorithm needs a predefined map; hence, it demands for accurate sensors for precise map building (Ibragimov IZ, Afanasyev IM in Comparison of ROS-based Visual SLAM methods in homogeneous indoor environment, IEEE, 2017). To overcome these challenges, the approach of reinforcement learning can be used (Appl Sci 10:8386, 2020). In reinforcement learning, the robot acts as an agent and performs navigation based on the received rewards. Here, our goal is to implement a DQN-based navigation algorithm on real 2WD robot for performing a mapless navigation. The sensor data such as data from LIDAR, IMU and motor encoders is fused together to perform the localization and navigation using the DQN algorithm. The results of navigation are compared against the benchmark SLAM algorithms like GMapping, Hector SLAM and Cartographer.