Failure Handling In a Planning Framework

When the real outcomes of actions are not completely represented, a planner may not be able to construct a valid plan even if there exists one. This research focuses on generic domain update and reasoning methods to construct alternative plans against real-time execution failures which are detected either during runtime or earlier by a plan simulation process. Based on the updated domain representations, a new executable plan is constructed even when the outcomes of existing operators are not completely known in advance or valid plans are not possible with the existing representation of the domain. TLPlan, a forward chaining temporal planner originally proposed by Bacchus and Ady (2001), is used as the temporal planner in this system since it can construct plans for durative and concurrent multirobot actions. A search node in TLPlan includes the world state and the action applied along with the world clock that defines the start time of that action. Applying an action at a state means that this action is scheduled for the world clock of that state. The search proceeds by applying new actions or advancing the world clock toward finding a complete temporal plan. When the agent executes its plan, the execution may be ceased due to unforeseen cases. Under these circumstances, the planner may not be able to construct a valid plan. In order to enable the agent construct a new plan, failures should be handled by investigating the causes of failures if reasoning tools are available. Some of the earlier studies have investigated action execution failures in planning frameworks. Goebelbecker et al. (2010) deal with a problem that is quite similar to our problem. Their purpose is finding explanations for failures by generating excuses, and making the agent replan with the help of these excuses. Excuses are extracted by changing the initial state such that a valid plan can be constructed from there on. Beetz et al. (2010) propose a system to cope with plan execution flaws. Their system involves plan projection and transformation processes in order to detect behavior flaws and generate a new plan. Plan projection is achieved by simulating the plan execution in a physics-based simulator. In