Local short path generation for autonomous commercial vehicles

Abstract

In recent years, breakthroughs in artificial intelligence (AI) have drawn the attention to the subject from many fields including the automotive industry, where it could become a cornerstone in order to develop fully autonomous vehicles. In the automotive industry the applications for these AI-techniques varies from classification of a vehicle's surroundings to behavioral-re ex approaches that mimics human behaviour. In this master thesis, the capability to navigate a truck in mining environments using neural networks has been investigated, tested and verified in a simulated 3D environment. As input to the neural networks, Light Detection And Ranging (LIDAR) sensors in different configurations has been used. The main focus has been to create an algorithm that can create short paths at a high rate using limited computational power. Consequently, the networks has been tried on Raspberry Pi to prove their capability. Several approaches are proposed using both 2D LIDARs as well as 3D LIDARs. The developed networks are simple, does not require high performance computational units and are able to make decisions at intersections according to a global planner. Apart from the developed networks, a tool-chain for collection of training data, network training and testing in simulated environment is described in detail in the report.