(A) path planning RL based on TD3 network embedded with AIRNN model

Abstract

We propose in this thesis a novel Reinforcement Learning based approach implemented in mobile robots to perform path planning task autonomously. The approach is titled “TD3 network embedded with Attention-based Identity Recurrent Neuron Network (AIRNN)”, since its structure is using TD3 network as a baseline model added to it a vanilla RNN model as a memory-based model denoted as IRNN and supported by attention mechanism. This model was trained and tested in a simulation environment that only use 2D LiDAR data points, distance to the target and final orientation as states and forward and angular velocities as actions. The main study focuses on investigating the feasibility of the model to find a path in an environment that has a target close to an obstacle without collision. We compared our model with TD3 network as a pure base structure and Attention-based Gated Recurrent Unit (AGRU) network as a competing structure that has similar structure to our model. The results showed that our model has surpassed the other models in terms of finding the shortest path with low computation time and ensuring the least collision occurrence.