Diagnosing Failures and Predicting Safe Runs in Robot Control

We present a formal framework for diagnosing failures and predicting safe runs of actions, given incomplete information on the initial state of affair, i.e. on the initial database providing a description of the domain before any action has taken place. Two aspects of uncertainty are formalized by two different notions of probability: uncertainty in the initial database and uncertainty during the execution of a course of actions. We introduce also a third concept of probability, which is obtained by combining the two previous notions. We call this new concept expected probability: it accounts for the probability of a sentence on the hypothesis that the sequence of actions needed to make it true might have failed. This new probability can be used to predict what would happen after a given course of actions; it leads to the possibility of comparing courses of actions and verifying which one is more safe. On the other hand, after a course of actions has been executed, and failures might have occurred, we give a methodology to diagnose what had really happened during the execution