Using source code to solve control problems

Abstract

Planning for long-horizon tasks in environments with non-discrete state spaces and dynamics with discontinuities remains a core challenge in robotics. In this setting, fully automatic search methods do not yet scale to many real-world problems of interest, and because of this, specialized planning algorithms (e.g., hierarchical planners) have been developed that leverage domain knowledge to organize the search for a successful plan. However, these specialized algorithms rely on representations tailored to specific problems and domains, which imposes additional workload. Recent work, however, has studied scalable techniques for finding concrete control inputs using a given control specification alone in the form of a logical formula, which reduces the burden on the user.

This thesis studies the application of program analysis techniques to the aforementioned planning problem, in conjunction with local formulae and hybrid search spaces in the style of hierarchical planners. Our observation is that the high-level structure of problem domains can often be coded into domain-specific simulators that model the high-level dynamics of the domain. This presents an opportunity to reuse that structure when describing the planning domain. We argue, this decreases the effort required to implement a planning system when a domain expert can relate domain knowledge to simulator source code. Thus, we design a planning system which can leverage simulator source code when describing a planning domain.