Development of an unmanned vehicle course control system based on reinforcement learning

At present, there is a growing development of autonomous transportation, driven by the need to improve road safety, reduce collisions, and enhance the efficiency of logistics operations. This trend is also influenced by increasing complexity in road conditions and challenges related to vehicle navigation and control, which make traditional control algorithms insufficient in terms of quality and effectiveness.

Aim. The objective of this research is to develop an intelligent system that enables an autonomous vehicle to independently control its course. The autonomous agent (a vehicle model) learns to navigate and follow a predefined trajectory using reinforcement learning through interaction with a simulation environment, based on the Actor-Critic method.

Materials and Methods. In this work, the Stable-Baselines 3 (SB3) library built on the PyTorch framework was used to implement and train the reinforcement learning model. The DonkeyCar simulator served as the training environment. To improve the speed and efficiency of training, a denoising autoencoder algorithm was applied to extract the region of interest (ROI).

Results. A series of comparative experiments was conducted to evaluate the impact of various parameters on training efficiency – such as speed limits, steering angle constraints, allowable deviation width from the lane center, movement continuity, discount factor, and frame rendering rate.

Conclusion. The results of the study demonstrate the potential of reinforcement learning in the field of autonomous transport,  while  also  highlighting  the  need  for  further  training  on  real-world  data, the prospects for scaling the approach to different classes of vehicles, and limitations related to computational resources and the need for safe behavior verification.