Appears in Working Notes of AAAI Spring Symposium on Planning with Incomplete Information for Robot Problems

Planning with Incomplete Information for Robot Problems Stanford, CA, March 1996, pp. 71-73 Planning for Murphy's Law: Uncertainty in CIRCA David J. Musliner & Kurt D. Krebsbach Automated Reasoning Group Honeywell Technology Center 3660 Technology Drive Minneapolis, MN 55418 fmusliner,[email protected] Abstract Anything bad that can happen will happen. Plan accordingly. Introduction Many planning and control systems attempt to represent their \degree of uncertainty" and incomplete information using quantitative measures (e.g., probabilities) or other techniques. Interestingly, our work has shown that this level of detail is relatively unimportant in mission-critical domains. When certain types of events and conditions are considered catastrophic (and hence wholly unacceptable), degrees of uncertainty become a moot point: Murphy's Law must be observed and planned for. Any possible way of reaching a catastrophic failure condition must be planned for and eliminated in order to provide guarantees of safe system performance. The CIRCA architecture was designed to provide predictable real-time performance and guaranteed system safety in mission-critical domains. The current implementation of CIRCA embodies a wide variety of techniques for dealing with uncertainty and incomplete information in several forms. In particular, CIRCA builds plans that can tolerate: Uncertainty in the timing characteristics of actions and exogenous processes. Uncertainty in action consequences. Uncertainty in future goals. Uncertainty in system state. Uncertainty in initial conditions. In addition, the CIRCA model of planning explicitly de nes a notion of \completeness" for plans used in interleaved planning and execution. This de nition justi es CIRCA's claims to real-time performance guarantees. In the following sections we provide a brief overview of the CIRCA architecture followed by additional details on how each of these types of incomplete information is handled in the CIRCA approach. sensor data Environment