Synthesis of Shared Autonomy Policies With Temporal Logic Specifications

Abstract

We propose a synthesis method of switching control policies for a class of shared autonomy systems in which control authority is held by either a human operator or an autonomous controller based on the state of the overall system. The objective is to optimize the system performance measured by the probability of satisfying a system specification in linear temporal logic, while ensuring a reasonable workload for the human operator. The synthesis method builds upon the construction of an abstract model for the given shared autonomy system from a set of components modeled by Markov decision processes, which are capable of capturing the uncertainty in the operator's performance and response to switching control signals under his different cognitive and physiological states. A cost function is then introduced to quantify the human operator's workload. In order to establish quantitative trade-offs between the operator's effort and the system performance, we propose a two-stage policy synthesis algorithm for generating Pareto-optimal switching control policies.