Continuous Motion Planning for Mobile Robots Using Fuzzy Deep Reinforcement Learning

Autonomous navigation for mobile robots has found a promising solution with deep reinforcement learning (DRL). This method stands out for its efficiency, particularly because it can operate without requiring extensive labeled datasets during the training process. Current DRL-based motion planning approaches can be categorized into two groups: DRL with discrete action spaces and DRL with continuous action spaces. The former aims to select the best action from a set of predefined actions but suffers from high computational costs. The latter computes actions directly from the continuous action space but may encounter local optima, leading to motion failures. To overcome these challenges, we introduce a novel approach, integrating Deep Deterministic Policy Gradient (DDPG) with fuzzy logic, for the motion of mobile robots. The Actor and Critic networks in our DDPG framework feature an architecture incorporating a Long Short-Term Memory (LSTM) network to handle varying numbers of obstacles in the environment. To address local optima in continuous DRL, we utilize fuzzification to discretize continuous actions into a small number of fuzzy sets and defuzzification to generate continuous actions. Consequently, the Actor generates fuzzy membership degrees, and the defuzzification process maps these degrees into continuous actions. In this way, our method leverages low-dimensional discrete DRL to ensure task completion. We conduct comprehensive experiments to evaluate the performance of our method. Compared with other DRL-based methods, the results show that our method can generate collision-free and smooth motion trajectories in real time and improve motion planning performance significantly.