Abstract
The requirement for autonomous robots to exhibit higher-level cognitive skills by planning and adapting in an ever changing environment is indeed a great challenge for the AI community. The goal of this research is to demonstrate how a physical robot can be capable of adapting its symbolic knowledge of the environment, by using experiences in robot action execution to drive knowledge refinement, and hence to improve the success rate of the task plans the robot creates. This thesis proposes a method for refining domain knowledge, encoded in the PDDL language, based on a novel method in reasoning using anomaly detection techniques.The ability for a robot to refine its knowledge of the environment is important when aiming for long term autonomy, enabling robots to adapt to changing situations, and to produce more robust task planning. The approach used in this thesis is to extend a high-level planning platform (ROSPlan) to help create more robust planning systems by using theory refinement techniques to improve the knowledge on which intelligent robot behaviour is based. The platform created combines planning, reasoning and incremental learning based on experience. Refinement involves reasoning over action execution failure in order to determine the cause of the failure (the anomaly), then using that cause to drive changes in domain knowledge, so that future abstract plans that are synthesised by the robot from the refined knowledge will lead to execution success.
This thesis reports on the construction of a software architecture within a real NAO robot, in order to support the software layers that eventually allow the robot to refine it’s knowledge. It details the algorithms used for determining the cause of failure, and for updating the faulty knowledge. It includes empirical testing and evaluation, using both real and simulated robot behaviour within a simple kitchen scenario. The results demonstrate that the implementation utilises its successes and failures in robot action execution to do knowledge refinement. Further the results show that, in the kitchen scenario, the refined knowledge leads to the synthesis of task plans which demonstrate decreasing rate of failure over time as faulty knowledge is removed or adjusted.
| Date of Award | 4 Nov 2022 |
|---|---|
| Original language | English |
| Supervisor | Lee McCluskey (Main Supervisor) & David Peebles (Co-Supervisor) |