Steering Angle Prediction by a Deep Neural Network and its Domain Adaption Ability

Abstract: The goal of this thesis is to design an artificial neural network for self-driving vehicles in regards to steering the vehicle. The performance of the networks is evaluated on a simulated race track in addition to more conventional metrics such as cost/loss. In the main part of the project the networks were trained on data from real-life driving and validated/tested on simulated data, which is an example of domain adaption. The simulated data were from the same conditions as in the simulated race track. Two main designs were evaluated, one based on the design proposed by NVIDIA and one based on the idea of multi-task learning where an autoencoder was trained simultaneously with a steering angle predictor. The resulting network based on the multi-task learning and solely trained on real-life driving data managed to make it around the entire simulated track without driving off the road. The network based on the NVIDIA design on the other hand only managed to stay on the road for a short period of time under the same conditions. The results indicate that the multi-task based design is better at domain adaption than the one based on NVIDIA's design and incites for further research within the area. Also some additional tests were conducted with real-life data for both training and validation. The results from this part of the thesis were ambiguous with respect to the domain adaption ability.