Behavior-Based Control for Autonomous Underwater Exploration

Transcription

1 Behavior-Based Control for Autonomous Underwater Exploration Julio Rosenblatt, Stefan Willams, Hugh Durrant-Whyte Australian Centre for Field Robotics University of Sydney, NSW 2006, Australia Abstract We present a system for behavior-based control of an autonomous underwater vehicle for the purpose of inspection of coral reefs, a task currently performed by divers holding a video camera while following a rope. Using sonar and vision-based approaches, behaviors have been developed for guiding the robot along its intended course, for maintaining a constant height above the sea floor, and for avoiding obstacles. A task-level controller selects which behaviors should be active according to user-defined plans and in response to system failures. Behavior arbitration has been implemented using both fuzzy logic and utility fusion. Initial experiments have been conducted in a natural coastal inlet, and the system is to be soon demonstrated in the coral reef environment. 1 Introduction Autonomous Underwater Vehicles have the potential to play a substantial role in ecology management, geophysical surveying, offshore oil and gas exploration and well maintenance, undersea mineral exploration and mining, and in surveillance and defence. For example, a survey project for reef management, carried out by the Australian Institute of Marine Sciences, is designed to provide long-term quantitative data about corals, algae and marine life over the extent of the Great Barrier Reef. This data is for studies of abundance and population change in selected organisms on a large geographic scale. Currently, visual transect information is recorded using underwater video cameras held by a diver following a rope, as shown in Figure 1. Figure 1: Diver taking visual transect of the reef This project presents an interesting opportunity for the introduction of an autonomous underwater vehicle to perform the task. However, the subsea environment is particularly unstructured and dynamic, the degrees of freedom in the control and estimation problem are greater than on land, and there exists no reliable positioning data. To function effectively in such conditions, an autonomous system must be responsive to its environment, proceeding in a data-driven manner, as well as goal-oriented, taking into account the higher level goals of the system. In order to achieve this desired symbiosis of deliberative and reactive elements, the Distributed Architecture for Mobile Navigation (DAMN) consists of a group of distributed modules communicating with a centralized command arbiter [9]. In order to avoid the bottlenecks and brittleness of centralized systems, DAMN is composed of specialized task-achieving modules, or behaviors, that operate independently and asynchronously. A behavior encapsulates the perception, planning and task execution capabilities necessary to achieve one specific aspect of robot control, and receives only that data specifically required for that task [2]. The reef surveillance task, as it is currently defined, consists primarily of following an assigned path while maintaining a fixed altitude above the reef and avoiding collisions. Independent behaviors and arbiters, using decoupled controllers, have been developed as a modular means of accomplishing these various sub-tasks [15]. For example, two behaviors have been developed for the path following aspect of the task; the first behavior uses video input to track a rope laid out along the coral, while the second behavior uses sonar to detect passive beacons. The DAMN arbiters are then responsible for combining the behaviors votes to generate controller commands. Fuzzy logic arbiters are currently used for control of the vehicle; another set of behaviors and arbiters that perform utility fusion [10] are under development. In both cases, the distributed, asynchronous behaviors provide real-time responsiveness to the environment, while the centralized command arbitration provides a framework capable of producing coherent behavior. A task-level controller selects which behaviors should be active according to user-defined plans and in response to system failures, based on knowledge of which behaviors are most appropriate in a given situation.

2 2 Oberon Submersible Vehicle We have constructed a simple low-cost underwater robot, shown in Figure 2, as a test-bed for experimental work in autonomous undersea navigation. The vehicle, named Oberon, has five thrusters; three in the vertical direction and two directed horizontally, giving it five independent degrees of freedom that enable the vehicle to maneuver in any direction except for purely lateral motion. The vehicle is equipped with two sonars and a color CCD camera, together with bathyometric depth sensors, a fiber optic gyroscope, and a compass. Figure 2: Oberon submersible vehicle The submersible vehicle is tethered to a ground base station which provides power, inter-processor communications, video transmission, and serial lines. The on board computer is directly interfaced to the vehicle hardware and is used to control the motion of the robot and to acquire sensor data. This data is collated and sent to the off-board computers via a tether for further processing and for the user interface. 3 Behavior-Based Control Architecture The Distributed Architecture for Mobile Navigation consists of a group of distributed behaviors communicating with centralized command arbiters, as shown in Figure 3. The behaviors use domain and procedural knowledge to vote for desirable actions and against objectionable ones. The arbiter is responsible for combining the behaviors votes to generate actions which are then sent to the vehicle controller. A task-level controller selects which behaviors should be active according to user-defined plans and in response to system failures, based on knowledge of which behaviors are most relevant and reliable in a given situation. One type of arbiter used for control of the vehicle employs fuzzy logic [17]; the behaviors send depth or heading commands as fuzzy sets. Another type of arbiter used performs utility fusion; instead of voting directly for actions, behaviors vote for the utility of possible outcomes and their associated uncertainty, and the arbiter determines which action will maximize expected utility. Each of these schemes has their own advantages and disadvantages, and both have been demonstrated to be effective for navigation of land-based vehicles [4][10]. Task-Level Controller activation status Avoid Obstacles Follow Targets Figure 3: DAMN consists of centralized arbitration of votes from distributed behaviors, activated by the task-level controller based on vehicle and behavior status. 3.1 Decoupled Control Control of an mobile robot in six dimensional space in an unstructured and dynamic environment would be a highly computationally intensive endeavor, even more so with the difficulties in controlling an underwater robot. We instead chose to decouple the control of motion in the horizontal plane from motion along the vertical axis. Thus, the individual design of controllers, behaviors, and arbiters for each subproblem is greatly simplified. This decoupling of the control problem is reasonable for the types of task we anticipate performing with underwater vehicles. In the reef surveying task, for example, we need to maintain a fixed height above the sea floor, independent of the path following behaviors that operate in the horizontal plane. In addition, we are not interested in explicitly controlling the vehicle s roll or pitch, merely minimizing them. Thus, there is one PID controller for the vertical thrusters which servos on depth, and another PID controller for the horizontal thrusters which governs yaw and speed. These controllers can then be designed, tuned, and operated independently of each other. There are also two arbiters in the system, each one corresponding to a separate controller. Likewise, behaviors are separated into those involved in movement along the Z-axis and those involved in movement in the X-Y plane, as suggested in Figure Behaviors DAMN Arbiters Maintain Altitude Minimum Depth Follow Rope DAMN BEHAVIORS votes commands data Vehicle In order to perform the task of surveying a coral reef transect, there are at least three behaviors that must be part of the system: following the transect path, maintaining a certain height above the coral, and avoiding collisions.

3 Because there is no global positioning data and only highly inaccurate internal data available to the robot, we attempt only to maintain position relative to sensed features in the environment. Two methods currently being used are vision-based navigation using a rope laid out along the coral and sonar-based navigation using passive beacons. They each have different failure modes and can be used in a complementary manner. Research into the simultaneous localization and mapping is also being conducted [5][14]and will be incorporated in the future. Maintain Altitude. This behavior provides the ability to maintain the robot at a certain altitude above the sea floor. In the context of this project, the robot must be able to detect the coral reef and maintain a relatively constant height above it in order to videotape the coral specimens and of course to avoid colliding with them. A sonar is used to periodically determine the altitude of the robot; the difference between the desired and actual altitudes then determines the amount by which this behavior would like to change the depth. This difference is subtracted from the current measured depth to set a new desired depth for this behavior, as illustrated in Figure 4. depth altitude detect obstacles measure altitude Figure 4: Sonar-based behaviors: Maintain Altitude uses altitude measurement to set desired depth; Avoid Collisions detects forward obstacles. Avoid Collisions. When the sonar is pointed forward, its returns are processed to determine if there are any obstacles that may lie ahead. For the fuzzy arbiter, the behavior would vote against the heading in which the obstacle lies; for the utility fusion arbiter it would indicate a negative utility at the computed obstacle location, along with uncertainty estimates. Follow Rope. The vision-based approach uses camera images in order to detect the survey line by color. A target color is specified, and a distance metric for image pixels is defined as the sum of the difference between a given pixel s RGB values and the target RGB values, normalized by pixel intensity: R D t R p + G t G p + B t B p p = R p + G p + B p (1) A random sample of pixels are then chosen from the image, and a dynamic programming search [6] is begun from each pixel to find a least cost path from the top to the bottom of the image, using the color distance metric defined above, as shown in Figure 5a. Figure 5: Path following behaviors: a) Follow Rope performs visual tracking; b) Follow Targets uses sonar to detect two targets in natural terrain. Once the line has been found in the image, a pursuit algorithm is used to determine which orientation would best bring the robot onto the line and sends that as a vote to the fuzzy arbiter. The location of the line itself can simply be sent to the utility fusion arbiter as a positive utility. Follow Targets. At present, sonar targets are introduced into the environment at the two ends of the transect to be followed. These targets present large, easily detectable features for the sonar, as indicated in Figure 5b. By maintaining a fix on two of these targets, the robot can find its current position and orientation relative to the line joining the two targets. Multiple targets can be used, depending on the roughness of the terrain and the length of the transect. The redundant information will also allow a better estimate of the line s position and orientation, thus improving line tracking. Using the range and bearing information to both of the targets to define the line to be followed, errors can be calculated in distance and orientation between the principle axis of the sub and the line. These two values can then be used to compute a new desired heading that will keep the sub on its desired path [16]; this is used to define the vote sent to the fuzzy arbiter. Again, for the utility fusion arbiter, the line is sent as a positive utility. 3.3 Fuzzy Logic Arbiter Fuzzy logic has found many uses for mobile robot control, including command fusion systems (see [12] for a survey). For the DAMN fuzzy logic arbiter, each behavior must determine the desired action, in this case either a heading or a depth, and vote for a fuzzy membership set with that action at the peak. Command defuzzification is Reef Targets

4 achieved using the maximum criterion to take the peak value; other defuzzification strategies, such as center of mass, assume a unimodal function and in general this averaging of inputs will select inappropriate commands. The fuzzy votes may take the form of a triangle or rectangle, as illustrated in Figure 6, or of a trapezoid. The figure shows votes from two behaviors. The first is the Move To behavior, which simply steers the vehicle to a specific set of coordinates; it is voting with a triangle whose apex is at 55 o. The second behavior is Avoid Collisions, which has a weight of 5; it has detected an obstacle ahead and is voting equally against all headings between 50 o and 70 o. The fuzzy sum of all votes is shown as a dashed line, and the output direction determined by the arbiter (40 o ) is shown as a dotted line. The current direction (60 o ) is shown in the figure as a solid black line. Figure 6: Fuzzy Arbiter 3.4 Utility Fusion Arbiter Utility Fusion is a novel means of action selection in which behaviors determine the utility of possible world states and their associated probabilities. These are collected by the arbiter and candidate actions are evaluated based on that information [10]. Utility theory provides a framework for defining votes and dealing with uncertainty. If assign a utility measure U(c) for each possible consequence of an action a, then the expected utility U(a) is: Ua ( ) = Uc () Pcae (, ) c where P(c a,e) is the probability that consequence c will occur given that we observe evidence e and take action a [7]. Thus, if behaviors provide these utilities and probabilities to an arbiter, it can then apply the Maximum (2) Expected Utility (MEU) criterion to select a Paretooptimal action based on all current information. [8]. Utility fusion does not create a world model as sensor fusion systems do. The information combined and stored by the utility fusion arbiter does not represent sensed features of the world, but rather the desirability of being in a particular state according to some criterion defined within the behavior. The processing of sensory data is still distributed among behaviors, so the bottlenecks and brittleness associated with sensor fusion are avoided. Unlike command fusion systems, the utility fusion arbiter does not simply select among or combine actions proposed by behaviors. Instead, behaviors indicate the utility of possible world states, together with estimates of uncertainty; the arbiter maintains a map of these utilities, and evaluates candidate actions within it. A more complete overview of various types of architectures and their corresponding advantages and disadvantages can be found in [11]. Through the use of utility theory, uncertainty within the system is explicitly represented and reasoned about within the decision-making processes. Utility theory teases apart the value of the consequence of an action from the probability that the consequence will occur and provides a Bayesian framework for reasoning about uncertainty [1]. This utility fusion approach deals explicitly with uncertainty, thus providing better defined vote semantics, and the utility map allows the central arbiter to use system models when evaluating actions in order to improve performance and stability. In the example shown in Figure 7, the two negative utilities correspond to obstacles detected by the Avoid Obstacles behavior; a line of positive utility is placed between the two sonar targets by the Follow Targets behavior, and the Follow Rope behavior has submitted another line of positive utility corresponding to the rope found in the video image. The arrows emanating from the vehicle suggest the candidate actions being evaluated by the arbiter. The one which has maximum expected utility is selected as the current commanded action. Visual Line Sonar Line Obstacles Figure 7: Utility Fusion Arbitration Sonar Targets

5 3.5 Task-Level Controller A task-level controller selects which behaviors should be active according to user-defined plans and in response to system failures, based on knowledge of which behaviors are most appropriate in a given situation. The system programmer defines schemas, parameterized planning constructs that indicate under which circumstances a behavior should run and how various success and failure signals sent by the behavior should be responded to. Schemas may be invoked within other schemas, thus creating a task tree whose root is the overall mission plan and whose leaves are the behaviors to be executed, as demonstrated in Figure 8. Each schema has a set of preconditions. A schema or behavior being invoked by a higher level schema becomes active when its preconditions are all true. In Figure 8, the Main schema has activated the Monitor Systems and Avoid Collisions behaviors, as well as the Survey Coral sub-schema which in turn enables the Record Coral, Follow Rope and Maintain Height behaviors, which will become active when their preconditions (e.g., the rope is visible) are true. At any given moment, the system s state of execution is defined by a hierarchy of active schemas and behaviors, the set of goals that they are trying to satisfy, and the values of various state variables. Thus, unlike TCA [13], the hierarchy is not strictly top-down and the behaviors do not operate in a fixed sequence, but rather are active only when all pre-conditions are satisfied. In addition, if one behavior or schema fails to produce the desired result, an alternative method will automatically be selected if available. This task-level controller is somewhat similar to he proposed continuous process extension to RAPS [3], with further capabilities and a more structured syntax. invoking some children and causing others to halt. If this schema is unable to handle the signal, it can instead pass the signal up to its parent, which then handles the exception in the same manner. A schema can indicate whether its children should be run in sequence or in parallel. If they are in sequence, then the behaviors and schemas will wait for the previous one to be successful before executing. Those in parallel will execute simultaneously until one or more of the children (as specified) signal completion. 4 Experimental Results This section presents some preliminary results from deployment of the vehicle in a natural terrain environment along Sydney s coastline. The first behavior developed was Maintain Altitude, which keeps the vehicle at a fixed standoff distance from the ocean floor. In the experimental results shown in Figure 9, the desired altitude was 1.5m, which was maintained within a standard deviation of 0.2m, as can be seen in the first plot if altitude vs. time. This is despite a rapidly changing bottom profile, as can be seen in the second plot of depth vs. time. Main Monitor Systems Record Coral Avoid Collisions Follow Rope Figure 8: Task tree. Survey Coral Maintain Height A behavior or schema signals success when it has successfully achieved its goal, otherwise it throws an exception back up to the calling schema when it has failed. This signal is caught by the parent, which can then assert or retract a predicate s truth value, effectively Figure 9: Results from Maintain Altitude behavior: a) plot of altitude vs. time shows that the desired altitude of 1.5m was maintained within 0.2m; b) plot of depth vs. time shows that the vehicle was continuously changing depth to match the profile of the sea floor. Executing the Follow Targets behavior resulted in the path shown in Figure 10 as a series of covariance ellipses representing 95% certainty in vehicle location. The vehicle was deployed in a natural coastal inlet with sonar targets placed in a line at 10m intervals. The areas on either side of the vehicle are reef walls detected and mapped from sonar data.

6 [5] Newman, P., and Durrant-Whyte, H. Toward Terrain- Aided Navigation of a Subsea Vehicle. In Proc. of International Conf. on Field and Service Robotics, Canberra, Australia, December [6] Nilsson, N. Principles of Artificial Intelligence, Tioga Pub. Co., Palo Alto, Calif., 1980 [7] Pearl, J. Probabilistic Reasoning in Intelligent Systems: Networks of Plausible Inference. Morgan Kaufmann Publishers, 1988 Figure 10: Path of robot executing Follow Targets behavior is shown as series of ellipses connected by a line, within a map created from sonar data. 5 Conclusion We have developed a behavior-based system for control of an autonomous underwater vehicle performing a survey of coral reefs. Implemented behaviors provide the ability to avoid collisions, maintain a proper standoff distance, and following the transect either using a rope with video or targets with sonar. Command fusion is performed using a fuzzy logic arbiter, and a utility fusion system is under development. A task-level controller selects which behaviors should be active according to user-defined plans and current environmental conditions. Initial experiments conducted in a natural coastal inlet have yielded promising early results, and the system is to be soon demonstrated in the coral reef environment. Acknowledgments The authors would like to thank the other members of the AUV project team at the ACFR, especially Ian Nowland, as well as Dr. Hugues Talbot of the CMIS CSIRO, who assisted in developing the rope detection algorithm. References [1] Berger, J. Statistical Decision Theory and Bayesian Analysis, 2nd ed. New York: Springer, [2] Brooks, R. A Robust Layered Control System for a Mobile Robot. Journal of Robotics and Automation, vol. RA-2, no. 1, pp , April [3] Firby, J. Task Networks for Controlling Continuous Processes. In Proc. of Second International Conf. on AI Planning Systems, Chicago, IL, June [4] Langer, D., Rosenblatt, J., and Hebert, M. A Behavior-Based System For Off-Road Navigation. Journal of Robotics and Automation, vol. 10, no. 6, pp , December [8] Pirjanian, P. The Notion of Optimality in Behavior-Based Robotics. To appear in Journal of Robotics and Autonomous Systems, [9] Rosenblatt, J. The Distributed Architecture for Mobile Navigation. Journal of Experimental and Theoretical Artificial Intelligence, vol. 9, no. 2/3, pp , April-September, 1997 [10] Rosenblatt, J. Utility Fusion: Map-Based Planning in a Behavior-Based System. Field and Service Robotics, Springer-Verlag, [11] Rosenblatt, J. and Hendler, J. Architectures for Mobile Robot Control. Advances in Computers, vol. 48, M. Zelkowitz, Ed., Academic Press, London, 1999 [12] Saffiotti, A. The Uses of Fuzzy Logic in Autonomous Robotics: a catalogue raisonne. Soft Computing 1(4): , Springer-Verlag, [13] Simmons, R. Structured Control for Autonomous Robots. Transactions on Robotics and Automation, 10:1, February [14] Williams, S., Newman, P., Dissanayake, G., Durrant-Whyte, H. Autonomous Underwater Simultaneous Localisation and Map Building. International Conf. on Robotics and Automation 2000 April 24-28, 2000, San Francisco. [15] Williams, S., Newman, P., Majumder, S., Rosenblatt, J., Durrant-Whyte, H. Autonomous Transect Surveying of the Great Barrier Reef. In Proc. of Australian Conf. on Robotics and Automation (ACRA'99), Brisbane, QLD, March 30-April 1, [16] Yuh, J. Modeling and Control of Underwater Robotic Vehicles. Transactions on Systems, Man and Cybernetics, vol 20., No. 6 November, pages , 1990 [17] Zadeh, L. Outline of a New Approach to the Analysis of Complex Systems and Decision Processes. Transactions on Systems, Man and Cybernetics, Vol 3 No 1, January 1973.