A three-tier architecture has been widely adopted for the task of controlling behaviour on board robots. The architecture acts to separate high-level deliberation over the best actions to take, from the low-level reactive task of executing those actions in a real dynamic and uncertain world. An important aspect that supports the efficiency of this architecture is that each tier acts with an increasingly abstract model.This allows the deliberation layer to construct plans for complex objectives in large domains. However when it comes to execution, the assumptions made to produce these plans can often contradict the realities of the environment they are required to operate in. This has resulted in the need for intelligent plan execution systems, which can monitor the execution of a plan and correct it if it an unexpected outcome is encountered.An introspective approach is presented, which focuses on increasing the intelligence of a system's executive (the mediating layer between deliberation and execution). This is achieved through the use of introspective models that provide the executive with an understanding of how each of its actions should progress. An introspective execution framework is constructed around these models, extending their use from monitoring into control.This provides a method for the executive to identify anomalous behaviour within an action's execution, and intervene with a recovery action before it results in a system failure. Problems within an execution can then be overcome at an executive level, without the need to abandon an action or invoke higher-level systems.The intelligence of a system's executive can then be further enhanced through the implementation of a stochastic controller that sits above the introspective framework. This provides additional reasoning over the executable actions, however it also provides the facility to add an element of learning to the system. By incorporating the observed outcomes from each action into the controller the executive can learn estimations of their success, and use this to refine a plan into a more robust execution strategy.This provides the executive with an expectation how an execution should be progressing, and raises the level on introspection to plan execution.This work is concerned with the ability to execute plans reliably on board a mobile robot operating in a dynamic environment and has therefore been implemented and deployed on a robot. Through several trials it is demonstrated that the introspective models can be used to predict action failure during action execution, which supports intervention and recovery. It is also shown that as the learning approach updates the stochastic controller, the system's behaviour changes adapting to its knowledge of the environment.
|Date of Award||18 May 2018|
- University Of Strathclyde
|Supervisor||Maria Fox (Supervisor) & Derek Long (Supervisor)|