Provably-Correct Task Planning for Autonomous Outdoor Robots

Abstract

Autonomous outdoor robots should be able to accomplish complex tasks safely and reliably while considering constraints that arise from both the environment and the physical platform. Such tasks extend basic navigation capabilities to specify a sequence of events over time. For example, an autonomous aerial vehicle can be given a surveillance task with contingency plans while complying with rules in regulated airspace, or an autonomous ground robot may need to guarantee a given probability of success while searching for the quickest way to complete the mission. A promising approach for the automatic synthesis of trusted controllers for complex tasks is to employ techniques from formal methods. In formal methods, tasks are formally specified symbolically with temporal logic. The robot then synthesises a controller automatically to execute trusted behaviour that guarantees the satisfaction of specified tasks and regulations. However, a difficulty arises from the lack of expressivity, which means the constraints affecting outdoor robots cannot be specified naturally with temporal logic. The goal of this thesis is to extend the capabilities of formal methods to express the constraints that arise from outdoor applications and synthesise provably-correct controllers with trusted behaviours over time. This thesis focuses on two important types of constraints, resource and safety constraints, and presents three novel algorithms that express tasks with these constraints and synthesise controllers that satisfy the specification. Firstly, this thesis proposes an extension to probabilistic computation tree logic (PCTL) called resource threshold PCTL (RT-PCTL) that naturally defines the mission specification with continuous resource threshold constraints; furthermore, it synthesises an optimal control policy with respect to the probability of success. With RT-PCTL, a state with accumulated resource out of the specified bound is considered to be failed or saturated depending on the specification. The requirements on resource bounds are naturally encoded in the symbolic specification, followed by the automatic synthesis of an optimal controller with respect to the probability of success. Secondly, the thesis proposes an online algorithm called greedy Buchi algorithm (GBA) that reduces the synthesis problem size to avoid the scalability problem. A framework is then presented with realistic control dynamics and physical assumptions in the environment such as wind estimation and fuel constraints. The time and space complexity for the framework is polynomial in the size of the system state, which is efficient for online synthesis. Lastly, the thesis proposes a synthesis algorithm for an optimal controller with respect to completion time given the minimum safety constraints. The algorithm naturally balances between completion time and safety. This work proves an analytical relationship between the probability of success and the conditional completion time given the mission specification. The theoretical contributions in this thesis are validated through realistic simulation examples. This thesis identifies and solves two core problems that contribute to the overall vision of developing a theoretical basis for trusted behaviour in outdoor robots. These contributions serve as a foundation for further research in multi-constrained task planning where a number of different constraints are considered simultaneously within a single framework.