Behavior-Based Robotics

Behavior-based robotics (BBR) bridges the fields of artificial intelligence, engineering, and cognitive science. The behavior-based approach is a methodology for designing autonomous agents and robots; it is a type of INTELLIGENT AGENT ARCHITECTURE. Architectures supply structure and impose constraints on the way robot control problems are solved. The behavior-based methodology imposes a general, biologically inspired, bottom-up philosophy, allowing for a certain freedom of interpretation. Its goal is to develop methods for controlling artificial systems (usually physical robots, but also simulated robots and other autonomous software agents) and to use robotics to model and better understand biological systems (usually animals, ranging from insects to humans).

Behavior-based robotics controllers consist of a collection of behaviors that achieve and/or maintain goals. For example, "avoid-obstacles" maintains the goal of preventing collisions; "go-home" achieves the goal of reaching some home destination. Behaviors are implemented as control laws (sometimes similar to those used in CONTROL THEORY), either in software or hardware, as a processing element or a procedure. Each behavior can take inputs from the robot"s sensors (e.g., camera, ultrasound, infrared, tactile) and/or from other behaviors in the system, and send outputs to the robot"s effectors (e.g., wheels, grippers, arm, speech) and/or to other behaviors. Thus, a behavior-based controller is a structured network of interacting behaviors.

BBR is founded on subsumption architecture (Brooks 1986) and other work in reactive robotics (RR). RR achieves rapid real-time responses by embedding the robot"s controller in a collection of preprogrammed, concurrent condition-action rules with minimal internal state (e.g., "if bumped, stop," "if stopped, back up"; Brooks and Connell 1986; Agre and Chapman 1987). Subsumption architecture provides a layered approach to assembling reactive rules into complete control systems from the bottom up. Rules, and layers of rules, are added incrementally; lower layers can function independently of the higher ones, and higher ones utilize the outputs of the lower ones, but do not override them. For example, "avoid-collision" at the lowest level, and "move-to-light" at a higher level, when combined, result in a robust light-chasing behavior; the higher-level rule never overrides the lower-level one, thus guaranteeing collision avoidance.

While robust, such reactive systems are limited by their lack of internal state; they are incapable of using internal representations and learning new behaviors. Behavior-based systems overcome this limitation because their underlying unit of representation, behaviors, can store state. The way state is represented and distributed in BBR is one of the sources of its novelty. Information is not centralized or centrally manipulated; instead, various forms of distributed representations are used, ranging from static table structures and networks to active, procedural processes implemented within the behavior networks.

In contrast to RR and BBR, both of which are structured and developed bottom-up, PLANNING-based deliberative control systems are top-down, and require the agent/robot to perform a sequence of processing sense-plan-act steps (e.g., "combine the sensory data into a map of the world, then use the planner to find a path in the map, then send steps of the plan to the robot's wheels"; Giralt, Chatila, and Vaisset 1983; Moravec and Elfes 1985; Laird and Rosenbloom 1990). Hybrid systems attempt a compromise between bottom-up and top-down by employing a reactive system for low-level control and a planner for high-level decision making (Firby 1987; Georgeoff and Lansky 1987; Arkin 1989; Payton 1990; Connell 1991). Often called "three-layer architectures," they separate the control system into three communicating but independent parts: (i) the planner, (ii) the reactive system, and (iii) the intermediate module, which reconciles the different time-scales and representations used by the other two and any conflicts between their outputs.

Behavior-based systems typically do not employ such a hierarchical division but are instead integrated through a homogeneous distributed representation. Like hybrid systems, they also provide both low-level control and high-level deliberation; the latter is performed by one or more distributed representations that compute over the other behaviors, often directly utilizing low-level behaviors and their outputs. The resulting system, built from the bottom-up, does not divide into differently represented and independent components as in hybrid systems, but instead constitutes an integra-ted computational behavior network. The power, elegance, and complexity of behavior-based systems all stem from the ways their constituent behaviors are defined and used.

Consequently, the organizational methodology of behavior-based systems differs from other control methods in its approach to modularity, the way in which the system is organized and subdivided into modules. Behavior-based philosophy mandates that the behaviors be relatively simple, added to the system incrementally, and not executed in a serial fashion. Subsets of behaviors are executed concurrently so that the system can exploit parallelism, both in the speed of computation and in the resulting dynamics that arise within the system itself (from the interaction among the behaviors) and with the environment (from the interaction of the behaviors with the external world). Behaviors can be designed at a variety of abstraction levels. In general they are higher than the robot"s atomic actions (i.e., typically above "go-forward-by-a-small-increment," "turn-by-a-small-angle"), and they extend in time and space. Some implemented behaviors include: "go-home," "find-object," "get-recharged," "avoid-the-light," "aggregate-with-group," "pick-up-object," "find-landmark," etc. Because behaviors can be defined at different levels of abstraction and can represent various types of information, they are difficult to define precisely, but are also a rich medium for innovative interpretations.

Deciding what behavior to execute at a particular point in time is called behavior arbitration, and is one of the central design challenges of BBR. For simplicity, most implemented systems use a built-in, fixed priority for behaviors. More flexible solutions, which can be less computationally efficient and harder to analyze, are commonly based on computing some function of the behavior activation levels, such as a voting or activation spreading scheme (Maes 1989; Payton et al. 1992). Behavior-based systems are typically designed so the effects of the behaviors largely interact in the environment rather than internally through the system, taking advantage of the richness of interaction dynamics by exploiting the properties of SITUATEDNESS/EMBEDDEDNESS. These dynamics are sometimes called emergent behaviors because they emerge from the interactions and are not internally specified by the robot"s program. Therefore, the internal behavior structure of a behavior-based system need not necessarily mirror its externally manifested behavior. For example, a robot that flocks with other robots may not have a specific internal "flocking" behavior; instead, its interaction with the environment and other robots may result in flocking, although its only behaviors may be "avoid collisions," "stay close to the group," and "keep going" (Mataric´1997).

Behavior-based robots have demonstrated various standard robotic capabilities, including obstacle avoidance, navigation, terrain mapping, following, chasing/pursuit, object manipulation, task division and cooperation, and learning maps, navigation and walking. They have also demonstrated some novel applications like large-scale group behaviors, including flocking, foraging, and soccer playing, and modeling insect and even human behavior (Agha and Bekey 1997; Webb 1994; Asada et al. 1994; Brooks and Stein 1994). Application domains have included MOBILE ROBOTS, underwater vehicles, space robotics, as well as robots capable of MANIPULATION AND GRASPING, and some walking and running machines.

Variations and adaptations of MACHINE LEARNING,and in particular REINFORCEMENT LEARNING, have been effectively applied to behavior-based robots, which have demonstrated learning to walk (Maes and Brooks 1990), navigate (Mataric´1992; Millan 1994), communicate (Yanco and Stein 1993), divide tasks (Parker 1993; Mataric´1997), behave socially (Mataric´1994), and even identify opponents and score goals in robot soccer (Asada et al. 1994). Methods from ARTIFICIAL LIFE, EVOLUTIONARY COMPUTATION, GENETIC ALGORITHMS, FUZZY LOGIC, VISION AND LEARNING, MULTIAGENT SYSTEMS, and many others continue to be actively explored and applied to behavior-based robots as their role in animal modeling and practical applications continues to develop.

-- Maja J. Mataric

References

Agha, A., and G. Bekey. (1997). Phylogenetic and ontogenetic learning in a colony of interacting robots. Autonomous Robots 4(1).

Agre, P., and D. Chapman. (1987). Pengi: an implementation of a theory of activity. Proceedings, Sixth National Conference of the American Association for Artificial Intelligence Conference. Seattle, WA, pp. 268-272.

Arkin, R. (1989). Towards the unification of navigational planning and reactive control. Proceedings, American Association for Artificial Intelligence Spring Symposium on Robot Navigation, Palo Alto, CA, pp. 1-5.

Asada, M., E. Uchibe, S. Noda, S. Tawaratsumida, and K. Hosoda. (1994). Coordination of multiple behaviors acquired by a vision-based reinforcement learning. Proceedings, IEEE/RSJ/GI International Conference on Intelligent Robots and Systems, Munich, Germany.

Brooks, R. (1986). A robust layered control system for a mobile robot. IEEE Journal of Robotics and Automation RA-2 (April), pp. 14-23.

Brooks, R., and J. Connell. (1986). Asynchronous distributed control system for a mobile robot. Proceedings, SPIE Intelligent Control and Adaptive Systems, Cambridge, MA, pp. 77-84.

Brooks, R., and L. Stein. (1994). Building brains for bodies. Autonomous Robots 1(1):7-25.

Connell, J. (1991). SSS: a hybrid architecture applied to robot navigation. Proceedings, International Conference on Robotics and Automation, Nice, France, pp. 2719-2724.

Firby, J. (1987). An investigation into reactive planning in complex domains. Proceedings, Sixth National Conference of the American Association for Artificial Intelligence Conference, Seattle, WA, pp. 202-206.

Georgeoff, M., and A. Lansky. (1987). Reactive reasoning and planning. Proceedings, Sixth National Conference of the American Association for Artificial Intelligence Conference, Seattle, WA, pp. 677-682.

Giralt, G., R. Chatila, and M. Vaisset. (1983). An integrated navigation and motion control system for autonomous multisensory mobile robots. Proceedings, First International Symposium on Robotics Research, Cambridge, MA: MIT Press, pp. 191-214.

Laird, J., and P. Rosenbloom. (1990). An investigation into reactive planning in complex domains. Proceedings, Ninth National Conference of the American Association for Artificial Intelligence Conference, Cambridge, MA: MIT Press, pp. 1022-1029.

Maes, P. (1989). The dynamics of action selection. Proceedings, International Joint Conference on Artificial Intelligence, Detroit, MI, pp. 991-997.

Maes, P., and R. Brooks. (1990). Learning to coordinate behaviors. Proceedings, Ninth National Conference of the American Association for Artificial Intelligence Conference, Cambridge, MA: MIT Press, pp. 796-802.

Matari, M. (1992). Integration of representation into goal-driven behavior-based robots. IEEE Transactions on Robotics and Automation 8(3):304-312.

Matari, M. (1994). Learning to behave socially. In D. Cliff, P. Husbands, J-A. Meyer, and S. Wilson, Eds., Proceedings, From Animals to Animats 3, Third International Conference on Simulation of Adaptive Behavior. Cambridge, MA: MIT Press, pp. 453-462.

Matari, M. (1997). Reinforcement learning in the multi-robot domain. Autonomous Robots 4(1):73-83.

Millan, J. (1994). Learning reactive sequences from basic reflexes. In D. Cliff, P. Husbands, J-A. Meyer, and S. Wilson, Eds., Proceedings, From Animals to Animats 3, Third International Conference on Simulation of Adaptive Behavior. Cambridge, MA: MIT Press, pp. 266-274.

Moravec, H., and A. Elfes. (1985). High resolution maps from wide angle sonar. Proceedings, IEEE International Conference on Robotics and Automation, St. Louis, MO.

Parker, L. (1993). Learning in cooperative robot teams. Proceedings, International Joint Conference on Artificial Intelligence, Workshop on Dynamically Interacting Robots, Chambery, France, pp. 12-23.

Payton, D. (1990). Internalized plans: a representation for action resources. In P. Maes, Ed., Robotics and Autonomous Systems, Special Issue on Designing Autonomous Agents: Theory and Practice from Biology to Engineering and Back 6 (1-2): 89 - 104.

Payton, D., D. Keirsey, D. Kimble, J. Krozel, and K. Rosenblatt. (1992). Do whatever works: a robust approach to fault-tolerant autonomous control. Journal of Applied Intelligence 3:226-249.

Webb, B. (1994). Robotic experiments in cricket phonotaxis. Proceedings of the Third International Conference on the Simulation of Adaptive Behavior. Cambridge, MA: MIT Press.

Yanco, H., and L. Stein. (1993). An adaptive communication protocol for cooperating mobile robots. In D. Cliff, P. Husbands, J. A. Meyer, and S. Wilson, Eds., Proceedings, From Animals to Animats 3, Third International Conference on Simulation of Adaptive Behavior. Cambridge, MA: MIT Press, pp. 478-485.

Behavior-Based Robotics

References

Further Readings