Context Dependant Iterative Parameter Optimisation for Robust Robot Navigation.

Progress in autonomous mobile robotics has seen significant advances in the development of many algorithms for motion control and path planning. However, robust performance from these algorithms can often only be expected if the parameters controlling them are tuned specifically for the respective robot model, and optimised for specific scenarios in the environment the robot is working in. Such parameter tuning can, depending on the underlying algorithm, amount to a substantial combinatorial challenge, often rendering extensive manual tuning of these parameters intractable. In this paper, we present a framework that permits the use of different navigation actions and/or parameters depending on the spatial context of the navigation task. We consider the respective navigation algorithms themselves mostly as a "black box", and find suitable parameters by means of an iterative optimisation, improving for performance metrics in simulated environments. We present a genetic algorithm incorporated into the framework, and empirically show that the resulting parameter sets lead to substantial performance improvements in both simulated and real-world environments in the domain of agricultural robots.