Permissive planning: extending classical planning to uncertain task domains

Reviewer: D. Charles Hair

The problem of automated planning in real-world settings is often complicated by the presence of uncertainty. This problem has been the subject of a number of research efforts in the AI community and has particular relevance in robotics, where real-time decisions must be made. The authors describe their development of a new approach that uses the idea of permissive planning. The approach is intended to achieve a balance between two potentially conflicting goals. One goal is to reach decisions in computationally reasonable amounts of time, where reaching that goal can lead to incorrect decisions because of a failure to account for real-world uncertainty. The other goal is to achieve correct results by reasoning about uncertainty, where achieving correct results can lead to unacceptably slow computations. The permissive planning approach uses a compromise approach in which reasoning about uncertainty is done during a separate learning phase. The separation of the learning phase from the planning phase means that the planning can be done efficiently in real-time settings. The paper presents an in-depth discussion of the permissive planning approach that includes theoretical justifications for the approach and an analysis of its correctness. In order to demonstrate the real-world usefulness of permissive planning, the authors also describe empirical work that was done with a robotic system called GRASPER. The paper is clearly written, includes a good list of references, and is highly technical. It should be of interest to anyone working in areas involved with planning, reasoning with uncertainty, and machine learning.