Abstract
A fundamental problem in creating successful shared autonomy systems is enabling efficient specification of the problem for which an autonomous system can generate a solution. We present a general paradigm, Interactive Shared Solution Shaping (IS3), broadly applied to shared autonomous systems where a human can use their domain knowledge to interactively provide feedback during the autonomous planning process. We hypothesize that this interaction process can be optimized so that with minimal interaction, near-optimal solutions can be achieved. We examine this hypothesis in the space of resource-constrained mobile search and surveillance and show that without directly instructing a robot or complete communication of a believed target distribution, the human teammate is able to successfully shape the generation of an autonomous search route. This ability is demonstrated in three experiments that show (1) the IS3 approach can improve performance in that routes generated from interactions in general reduce the variance of the target detection performance, and increase overall target detection; (2) the entire IS3 autonomous route generation system's performance, including cost of interaction along with movement cost, experiences a tradeoff between performance vs. numbers of interactions that can be optimized; (3) the IS3 autonomous route generation system is able to perform within constraints by generating tours that stay under budget when executed by a real robot in a realistic field environment.