Probabilistic self-tuning approaches for enhancing performance of autonomous vehicles in changing terrains

Motion controllers usually require a tuning stage to ensure an acceptable performance of the vehicle during operation in challenging scenarios. However, such tuning stage is a time consuming process for the programmer and often is based on intuition or heuristic approaches. In addition, once tuned, the vehicle performance varies according to the nature of the terrain. In this work, we study the use of well-known probabilistic techniques for self-tuning trajectory tracking controllers for service units based on the idea of saving both vehicle's resources and human labour force time. The proposed strategies are based on Monte Carlo and Bayesian approaches to find the best set of gains to tune the controller both off-line and on-line, thus enhancing the controller performance in the presence of changing terrains. The approaches are implemented and validated on a skid-steer mini-loader vehicle usually used for mining purposes. Implementation details and both simulation and empirical results are included in this work, showing that when using our approaches, effort can be saved up to 30% and tracking errors reduced up to 75%.