This paper describes an implementation of the 3T robot architecture which has been under development for the last eightyears. The architecture uses three levels of abstraction and description languages whichare compatible between levels. The makeup of the architecture helps to coordinate planful activities with real-time behaviors 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 varietyofvery different robot systems using different processors, operating systems, effectors and sensor suites.

The degree to which a planner succeeds and meets response deadlines depends on the correctness and completeness of its models which describe events and actions that change the world state. It is often unrealistic to expect perfect models, so a planner must be able to detect and respond to states it had not planned to handle. In this paper, we characterize different classes of these "unhandled" states and describe planning algorithms to build tests for, and later respond to them. We have implemented these unhandled state detection and response algorithms in the Cooperative Intelligent Real-time Control Architecture (CIRCA), which combines an AI planner with a separate real-time system so that plans are built, scheduled, and then executed with real-time guarantees. Test results from flight simulation show the new algorithm enables a fully-automated aircraft to react appropriately to certain classes of unhandled states, averting failure and giving the aircraft a new chance to achieve its ...

In this paper we briefly describe a software architecture that integrates reactivity and deliberation. The architecture has three levels, or tiers. The bottom tier contains a set of situated, reactive skills. The second tier is a sequencer that can activate sets of skills to perform a task. The third tier is a planner that reasons in-depth about goals, resources and constraints. We characterize this architecture across several dimensions and then describe the lessons that we have learned from applying this architecture to several robotics projects. Introduction For several years we have investigated ways to combine deliberation and reactivity in robot control architectures to program robots to carry out tasks robustly in field environments (Bonasso 1991; Bonasso, Antonisse, & Slack 1992). We believe this integration is crucial. Not only must an agent be able to adjust to changes in a dynamic situation, but it must be able to synthesize plans, since the complexities of the real world m...

This paper describes the software framework that has been developed and that it is being used to build a control architecture for the navigation of a B21 mobile robot. The subsumption architecture and the multi-agent based controller architectures are analyzed and a software framework named SORGIN is presented as a structured programming tool that allows to deal with the complexity of mobile robotic systems. We show how a task that performs safe wandering with a privileged compass orientation has been defined using di#erent behavior bricks in the SORGIN framework.

We present a complete architecture for behavioral control of locomotion for both real and simulated agents and provide a design methodology for building the locomotion control systems that embody the architecture. A low-level locomotion engine controls an agent's actions directly based on intermediate-level reactive behaviors such as attraction and avoidance. High-level state machines schedule and control the reactive behaviors allowing for more "intelligent" decision processes, and an agent model provides a mechanism for varying locomotion according to agent state and personality attributes. In addition to providing specifications for a locomotion engine, we address the problem of selecting and organizing an appropriate set of behaviors. We present selection criteria and a method for partitioning the behaviors to aid in implementation. We discuss the challenges specific to human locomotion and explain how to...

This paper describes a multi-layered, intelligent agent software architecture, developed for mobile and undersea robot applications in the defense sector, and to provide tele-autonomy to space-based manipulator robots. The architecture has a deliberative layer which uses a state-based planner, a middle layer for sequencing partially ordered plans using robot skills, and a lower layer repertoire of continuous robot skills. The system has been shown to provide a higher level of human supervision that preserves safety while allowing for task level direction, reaction to out-ofnorm parameters, and human intervention at all levels of control. For this workshop, we hypothesize that the architecture is a useful framework in which to explore learning techniques. In particular, we outline techniques appropriate to learning within a given layer, techniques for migrating competences from higher to lower layers, and overall system adaptation from its interaction with the environment. Examples are reinforcement learning for tuning individual skills, case-based techniques to improve the re-planning capability of the deliberative layer, and chunking or explanation-based learning to migrate new strategies created by the planner into standard procedures for the sequencing level. Background and Motivation Since the late eighties we have investigated ways to combine deliberation and reactivity in robot control architectures [Sanborn et al 1989, Bonasso 91, & Bonasso et al 92], in order to program robots to carry out tasks robustly in field environments. Field environments are those in which events for which the robot has a response can occur unpredictably, and wherein the locations of objects and other agents is usually not known with certainty until the robot is carrying out the required task. A robot control software architecture, developed at MITRE is an outgrowth of several lines of situated reasoning research in robot intelligence [Firby 89, Gat 91, Connell

this paper. "Planned-for" states are those the planner has expanded. This set is divided into two parts: "handled" states from which failure is assured to be avoided and from which the goal can be reached, and "deadend" states from which failure is avoided but from which the goal cannot be reached using the current plan. All World States Modeled Planned-for "Handled" -- can reach goal Deadend Removed Imminent Failure World States Actually Reached Figure 5-1: World State Classification Diagram. 84 A variety of other states are modelable by the planner. Such states include those identified as reachable, but "removed" because attending to them along with the "planned-for" states exceeds system capabilities. Other modeled states include those that indicate "imminent failure;" if the system enters these states, it is likely to fail shortly thereafter. Note that some states might be both "removed" and "imminent-failure", as illustrated in Figure 5-1. Finally, some modeled states might not fall into any of these categories, such as the states the planner considered unreachable but that are not necessarily dangerous. As illustrated by the boldly outlined region in Figure 5-1, states actually reached may include any subclass. To assure safety, the set should only have elements in the "planned-for" region. When the set has elements outside this region, safety and performance depend on classifying an unplanned-for state if and when it is entered and responding appropriately. For this reason, we provide more detailed definitions of the most important classes. A "deadend" state (DS) results when a transition path leads from an initial state to a state that cannot reach the goal, as shown in Figure 5-2. The deadend state is safe because there is no transition to failure. However, ...

Complex tasks are usually described as high-level goals, leaving out the details on how to achieve them. However, to control a robot, the task must be described in terms of primitive commands for the robot. Having the robot move itself to and through an unknown, and possibly narrow, doorway is an example of such a task. We show how the transformation from high-level goals to primitive commands can be performed at execution time and we propose an architecture based on reconfigurable objects that contain domain knowledge and knowledge about the sensors and actuators available. We illustrate our approach using actual data from a real robot. 1 Introduction A large part of our civilized world is unstructured and dynamic. This fact has been one of the leading obstacles to mass-quantity introduction of robots into everyday life. The variability of the world makes it impractical to develop detailed plans of actions before execution. The environment might change and thus invalidate the plan. I...

A robot must have an internal representation of the local space it occupies to use for both navigation and obstacle localization. In addition, it must be possible to build and update the map in real-time so that it can be used in feedback control loops. A robot's notion of local space must bridge the gap between symbolic and continuous control. To satisfy both real-time constraints and the needs of high-level navigation and object recognition, the map building system must use a simple representation that can be computed quickly yet will support the construction of more involved maps over longer timescales. A complete system also requires control behaviors that can use the simple representation to drive the robot through its immediate surroundings in service of higher- level local navigation goals generated from the more detailed map. This paper describes a system based on building simple geometric occupancy maps from multiple sensors in real-time and using them for control. ...