Mobile Robot Navigation Using Twin Delayed Deep Deterministic Policy Gradient and Fuzzy Logic

In this paper, a novel approach is proposed for navigating a mobile robot by combining a Deep Reinforcement Learning Algorithm (Deep RL), Twin Delayed Deep Deterministic Policy Gradient (TD3), and fuzzy logic. This method navigates the robot with continuous actions that are linear and angular velocities, enabling it to operate in both known and unknown environments. The states and data for this navigation control system are obstacle distances, goal distance, and goal angle, allowing the robot to operate in varying conditions effectively. Using the trained TD3 model, the robot moves in a straight line toward the goal point, while fuzzy logic ensures it is avoiding obstacles along the way. The TD3 model is trained in the Turtlesim environment and then the combination of the trained model and fuzzy network is validated in the Gazebo environment using the TurtleBot3 - Burger simulation. The goal distance and goal angle are calculated from data provided by the robot's odometer and gyroscope sensors, respectively. Thereafter, this data is processed by the TD3 network. Obstacle data is received from a 360-degree LiDAR sensor attached to the robot. When the robot reaches an obstacle with the distance of 0.07 and less (in the normalized scale), the fuzzy rules are applied to process this data and send immediate actions to prevent collisions.