A General Programming Language for Unified Planning and Control

摘要

This paper presents a method for embedding predictive search techniques within a general purpose programming language. We focus on using this language to program the behavior of a real-time control system. Our goal is the ability to write complex programs that can be interpreted by both a real-time controller and an associated planner. We present an expressive action representation which captures the procedural complexities of practical control programs, yet still can be projected by a search-based planner. To support integration with the real-time controller, the planner can be interrupted after an arbitrary amount of computation. Whenever the planner is interrupted, it can provide useful advice to the controller. The system represents a unified approach since the planner and the controller share identical data structures and algorithms for interpreting a shared action representation. This unified representation facilitates very tight integration between the planner and the controller. Our system increases the procedural expressiveness of planning techniques and provides a powerful software abstraction that facilitates the use of search techniques by control system programmers. The paper sets the context for embedded planning within real-time control systems, and then presents an overview of our approach. We then provide a detailed description of how the system works using a running example, and discuss the results of experiments that used the same example. This is followed by an evaluation of our assumptions, contributions, limitations and future work. We close the paper by presenting a NASA application that will drive much of our future work.