Domain Compilation for Embedded Real-Time Planning

摘要

A recently conceived approach to automated real-time control of the actions of a robotic system enables an embedded real-time planning algorithm to develop plans that are more robust than they would otherwise be, without imposing an excessive computational burden. This approach occupies a middle ground between two prior approaches known in the art as the universal-plan and hybrid approaches. Ever since discovering the performance limitations of taking a sense-plan-act approach to controlling robots, the robotics community has endeavored to follow a behavior-based approach in which a behavior includes a rapid feedback loop between state estimation and motor control. Heretofore, system architectures following this approach have been based, variously, on algorithms that implement universal plans or algorithms that function as hybrids of planners and executives. In a typical universal-plan case, a set of behaviors is merged into the plan, but the system must be restricted to relatively small problem domains to avoid having to reason about too many states and represent them in the plan. In the hybrid approach, one implements actions as small sets of behaviors, each applicable to a limited set of circumstances. Each action is intended to bring the system to a subgoal state. A planning algorithm is used to string these actions together into a sequence to traverse the state space from an initial or current state to a goal state. The hybrid approach works well in a static environment, but it is inherently brittle in a dynamic environment because a failure can occur when the environment strays beyond the region of applicability of the current activity. In the present approach, a system can vary from the hybrid approach to the universal-plan approach, depending on a single integer parameter, denoted n, which can range from 1 to a maximum domain-dependent value of M. As illustrated in the figure, n = 1 represents the hybrid approach, in which each linked action covers a small part of the state space of the system. As n increases, the portion of state space associated with each action and its subgoal grows. When n reaches M, coverage extends over the full state space, so that the system contains a universal plan.