R. James Firby

Experiences with an architecture for intelligent, reactive agents

This paper describes an implementation of the 3T robot architecture which has been under development for the last eight years. The architecture uses three levels of abstraction and description languages which are compatible between levels. The makeup of the architecture helps to coordinate planful activities with real-time behaviours for dealing with dynamic environments. In recent years, other architectures have been created with similar attributes but two features distinguish the 3T architecture: (1) a variety of useful software tools have been created to help implement this architecture on multiple real robots; and (2) this architecture, or parts of it, have been implemented on a variety of very different robot systems using different processors, operating systems, effectors and sensor suites.

Hierarchical planning involving deadlines, travel time, and resources

This paper describes a planning architecture that supports a form of hierarchical planning well suited to applications involving deadlines, travel time, and resource considerations. The architecture is based upon a temporal database, a heuristic evaluator, and a decision procedure for refining partial plans. A partial plan consists of a set of tasks and constraints on their order, duration, and potential resource requirements. The temporal database records the partial plan that the planner is currently working on and computes certain consequences of that information to be used in proposing methods to further refine the plan. The heuristic evaluator examines the space of linearized extensions of a given partial plan in order to reject plans that fail to satisfy basic requirements (e.g., hard deadlines and resource limitations) and to estimate the utility of plans that meet these requirements. The information provided by the temporal database and the heuristic evaluator is combined using a decision procedure that determines how best to refine the current partial plan. Neither the temporal database nor the heuristic evaluator is complete and, without reasonably accurate information concerning the possible resource requirements of the tasks in a partial plan, there is a significant risk of missing solutions. A specification language that serves to encode expectations concerning the duration and resource requirements of tasks greatly reduces this risk, enabling useful evaluations of partial plans. Details of the specification language and examples illustrating how such expectations are exploited in decision making are provided.

Representing and Solving Temporal Planning Problems

AI planning research has generally been concerned with the advantages and disadvantages of representing temporal planning information as a partially ordered task network. Such a network represents a growing plan as a set of tasks that become more completely temporally ordered as the plan is elaborated. A partially ordered task network is an attractive plan representation because it maintains the flexibility to avoid or take advantage of unexpected task interactions by simply adding appropriate ordering constraints between existing tasks. Unfortunately, a partially ordered network has difficulty representing tasks with properties that depend on precisely what other tasks come before them. How can one keep track of a bank balance in the face of unordered withdrawals and deposits? This paper discusses work done at Yale by Dean and Miller that addresses both the problem of building and maintaining a partially ordered task network and the problem of reasoning about tasks that have order-dependent properties.