Mapless Navigation System for Agriculture Mobile Robot with Deep Reinforcement Learning

The rapid development of technology has facilitated the implementation of robots in agriculture, posing issues such as the creation of navigation systems for vast fields or plantations, where mapping is unfeasible due to variable topography. The work proposes to design a navigation system for robots that operates without maps utilizing Deep Neural Networks as the motion controller and to establish a Deep Reinforcement Learning platform for automated parameter optimization. Simulation models of the robot and plantation environment were developed in MATLAB, employing an actor-critic network that integrates policy-based and value-based approaches. The neural network architectures for the actor and critic networks were established, and training was performed in simulation, successfully enabling autonomous navigation to specified target locations. Simulation was executed in two distinct configurations, decomposed navigation and combined navigation. The results demonstrated that the mapless navigation system displayed robust performance, with the decomposed navigation technique proving to be 24% more efficient than the combined navigation technique regarding drastic position changes and 4.5% faster in terms of navigation length of time.