Replanning: a New Perspective

Replanning involves generating a new plan to fix execution failures. Past work has often characterized replanning procedurally as a special case of Plan Reuse: reuse the current plan to solve a new problem with altered initial and goal states. Just as in normal reuse, these replanning systems focus on minimally perturbing the plan while accounting for the alterations in the problem. There are two important limitations to this view of replanning. First, many failures cannot be represented as alterations to initial and goal states. Second, plan reuse is motivated by efficiency considerations during plan generation, while replanning attempts to minimize the residual execution costs of the partial execution. In this paper, we argue that replanning should rightly be seen as solving a new planning problem that not only captures general execution failures, but also the commitments incurred during the partial execution. Failures are modeled as finegrained modifications to operator descriptions, and commitments are modeled as soft constraints. From this general perspective, minimal perturbation planning can be understood as a crude heuristic for respecting commitments: if we keep the plan the same, then we are likely to respect the commitments