Characterizing and evaluating autonomous controllers

Résumé

Autonomy in robotics by means of Artificial Intelligence (AI) Planning & Scheduling (P&S) is a widely research area with great interest in applications such as exploration robots in hazardous or human unreachable areas. However, autonomous controllers for robotics are usually not well assessed. For instance, it is not easy to compare newer assets with previous works in the field. In this thesis we propose a framework, called On-Ground Autonomy Test Environment (OGATE), to support testing and assessment of autonomous controllers. It is supported on a methodology and a set of generally applicable and domain independent metrics to generate objective evaluations, and a software tool to enable automatic benchmarking processes. To demonstrate the effectiveness of the framework, we exploit two autonomous controllers based on different P&S paradigms. First, the Goal Oriented Autonomous Controller (GOAC) developed under an European Space Agency (ESA) contract. Second, the Model-Based Architecture (MOBAR) developed during this PhD that exploits a different paradigm for autonomy. Particularly, MOBAR is designed with the objective of testing different Planning Domain Definition Language (PDDL) based planners to achieve on-board autonomy. In this regard, we also introduce a new planner, Unified Path Planning and Task Planning Architecture (up2ta), that integrates a state of the art PDDL planner with path planning algorithms. The objective of up2ta is to produce efficient plans for robotics exploration missions. Regarding to the path planning algorithm, in this thesis we introduce two algorithms focused on mobile robots: S-Theta* that effectively reduces the heading changes of the path, and the 3D Accurate Navigation Algorithm (3Dana) that deals with Digital Terrain Models (DTMs) and traversability cost maps to produce safer and reachable paths in realistic environments. Given both controllers, OGATE has been successfully exploited to evaluate them, allowing to characterize relevant aspects about the integration between Planning & Execution (P&E) that are hardly to be assessed in other ways. Moreover, the results are reproducible and objective, enabling comparison among controllers with different P&S technologies and paradigms.