Issue No. 02 - April (1995 vol. 10)
DOI Bookmark: http://doi.ieeecomputersociety.org/10.1109/64.395355
<p>A first-place finisher illustrates the advantages of hierarchical, behavior-based control, with low-level behaviors ensuring craft survival and high-level behaviors performing tasks such as navigation and object location. </p> <p>The International Aerial Robotics Competition is an annual event sponsored by the Association for Unmanned Vehicle Systems (AUVS). The competition requires flying robots to locate and manipulate objects and transport them from one location to another. The robots must perform these tasks under hazardous conditions, without human guidance, and within a fixed time limit.</p> <p>Creating a flying robot with these capabilities presents many challenges. To achieve its goals, the robot must make control decisions based on imperfect sensory data, while adapting to unexpected situations such as gusts of wind or sensor failure. It must make these decisions in real time to maintain the craft's safety and ensure system survival.</p> <p>Teams of university students from around the world enter the competition. The University of Southern California Robotics Research Laboratory's Autonomous Flying Vehicle-I (AFV-I) finished first among the more than twenty schools entered in the fourth annual competition held on May 19, 1994. The University of Texas at Arlington and the Southern College of Technology finished in second and third places, respectively.</p> <p>The flying robots use a variety of control approaches. The AFV-I uses a behavior-based control architecture, which partitions the control problem into a set of loosely coupled computing modules. Each module, or behavior, is responsible for achieving a specific task. These behaviors interact to achieve the robot's overall goals. The modules are organized hierarchically, with low-level, reflexive behaviors responsible for craft survival and high-level behaviors responsible for tasks such as navigation and object location.</p> <p>A behavior-based approach has many advantages over traditional methods of controlling autonomous mobile robots. Traditional methods attack the control problem sequentially. That is, a robot first senses, perceives, and models its environment; then it plans and acts in its environment. Since the world is information rich, traditional methods are prone to information overload, which renders the robot incapable of functioning in real time, possibly with dire consequences. In addition, these methods assume that the robot can construct accurate, global world models from the incoming sensory information. A number of factors make this difficult, such as a rapidly changing world, limited computer processing power, and inaccurate, incomplete sensor models.</p> <p>In contrast, a behavior-based approach solves the problem in a parallel fashion. Each behavior, acting concurrently with other behaviors, extracts from the environment only the information required to complete a given task at a given time, avoiding information overload. This division of labor also eliminates the need for construction and maintenance of a global world model, further reducing the robot's computational load.</p> <p>Another advantage of the behavior-based approach is that it gives us the ability to create layers of increasingly complex behaviors. If necessary, higher-level behaviors can inhibit or modulate lower-level behaviors. Thus, we can incrementally build and test a robot control system with increasing capabilities, without losing low-level capabilities already created. This control approach has been explored by others, including Brooks in the subsumption architecture and Arkin in AuRA, an architecture for reactive navigation.</p> <p>However, the behavior-based approach has limitations. Interactions and possible couplings of behaviors, which may critically affect the craft's stability, are unknown a priori. For example, changing the RPM of the main rotor changes the craft's yawing tendency, demonstrating a coupling of the heading control and thrust control behaviors. Determining such couplings experimentally may be necessary. Since no models are available, this experimentation can be time consuming and potentially hazardous to the craft. (Thus, one of the strengths of the approach can also be a problem.) Attempts to expand system complexity worsen the coupling problem by increasing the number of behaviors and layers. To overcome this problem, we are planning research on methods of developing behavior-based models from performance data, including methods by which the robot can obtain relevant parameters by learning.</p>
G. A. Bekey, J. F. Montgomery and A. H. Fagg, "The USC AFV-I: A Behavior-Based Entry in the 1994 International Aerial Robotics Competition," in IEEE Intelligent Systems, vol. 10, no. , pp. 16-22, 1995.