1 Developments in AI Planning Research T. K. Satish Kumar Knowledge Systems Laboratory Stanford University tksk@ksl.stanford.edu Abstract: An intelligent agent should be able to decide upon a course of actions to achieve certain goals. A plan is such a course of actions and the planning problem is the problem of finding a plan for a given representation of actions in the world the agent operates in. In this paper, we provide a survey of the developments in AI planning research, with an emphasis on recent techniques. Introduction, Background and Early Work: An intelligent agent should be able to decide upon a course of actions to achieve certain goals. A plan is such a course of actions and the planning problem is the problem of finding a plan for a given representation of actions in the world the agent operates in. GPS (general problem solver) was the first planner to distinguish between general problem- solving knowledge and domain knowledge. It introduced means-end analysis in its search procedure. This approach tried to find a difference between the current object and the goal, and then used a lookup table to invoke an action that would minimize the difference. If the action could not be applied to the current object, GPS recursively tried to change the object into one that was appropriate. Green’s approach to planning was to formulate the problem as theorem proving. Green introduced a ‘state’ or ‘situation’ variable into each predicate. The goal condition then consisted of a formula with an existentially quantified state. The system' s job was to prove that the goal condition was satisfiable---that is, there existed some state in which the condition was true. A special function do(action, state) denoted the function mapping a state into the state resulting from taking an action. Green' s formulation also used frame assertions to state what propositions did not change as the result of an action. A serious drawback however was that a frame assertion was required for every relation unaffected by an action. To simplify the frame assertions of Green' s formulation, Kowalski' s approach introduced the ‘ Holds’ literal. This allowed the results of each action to be framed by a single axiom. The planning problem was proven to be undecidable when the action representation includes actions whose effects are a function of their situation. A subset of planning problems however can be handled well by the so-called STRIPS representation of actions. It is a way of coping with the frame problem when the world states are represented by a set of propositions. Under this representation, there are a set of operators or action schema, where each operator has (1) a set of parameters like quantified variables (e.g. ?r in ‘rocket ?r’), (2) a set of preconditions, and (3) a set of effects consisting of add effects and delete effects. The add effects and delete effects are a list of propositions that will be added and deleted respectively from the current world state’s list of propositions to get the new world-state’s list of propositions. Every instantiation of an operator