Evolution of high level motion control for autonomous ground vehicles

Abstract

Autonomous robotic exploration is the task of building models of an environment. This task requires robots to rapidly plan, re-plan and execute their motion trajectories using sensory data that is provisional, uncertain and noisy. To navigate successfully under these conditions, robots require carefully designed motion controller software to guide the robot safely, quickly, reliably and efficiently to intermediate exploration objectives. Conventionally, the basic design of a motion controller is derived from first principles using simplified models of motion control and then refined by hand in response to observed performance. While this approach works in simpler applications, it becomes more challenging and less effective as applications become more complex and the number of variables to consider increases. Moreover, changes in robot configuration and environment can entail costly redesign of the controller. As such, we argue that this manual approach will become increasingly impractical as our exploration tasks become more ambitious. In this thesis, we address the development of motion control using techniques from Evolutionary Computation (EC). Our approach is to view the motion control design as a search problem, that can be subject to automation. In this work we present a novel framework for evolving the core component of motion control based on a form of EC called Grammatical Evolution (GE). GE systematically refines populations of potential programatic solutions for a given problem, until an effective solution is found. In our approach, we use GE to search automatically for the best motion control for a given set of exploration tasks. GE allows the user to constrain the search space for programs using Backus-Naur Form (BNF) grammar specifications. We use these grammars to define the search space for controllers for each exploration application. We conducted four experiments to evaluate our proposed approach. Each experiment demonstrates the framework in different exploration configurations and different requirements. All of our experiments evolved controller code for unmanned ground vehicles (UGV’s). Our first experiment evolved numerical parameters for the control of small teams of UGV’s. Our second experiment evolved control for a single UGV to optimise exploration performance and energy consumption. Our third experiment evolved both the structure and parameters of the core control function. Our fourth experiment evolved the input factor selection and numerical constants for well-established navigation approach in progressively more realistic situations - culminating in deployment on real platforms. In each of our experiments we found that the automated search approach outperformed carefully designed handwritten control. Moreover the structure of the evolved equation helped to reveal the nature of the trade-offs inherent in the exploration task and what factors appear to be most relevant to informing effective control.