Strategies for Searching an Area with Semi-Autonomous Mobile Robots Robin R. Murphy and J. Jake Sprouse 1 Abstract This paper describes three search strategies for the semi-autonomous robotic search of an area, and how they can be implemented via reactive behaviors within a semi-autonomous control architecture. The operator performs semantic search, the robot is responsible for systematic search, and opportunistic search is done cooperatively. The robotic search behavior is a collection of three reactive behaviors (navigation, scanning, and examination) coordinated by a search controller. The paper reports on work in progress in implementing these search strategies and architecture on the Colorado School of Mines mobile robot, Clementine. Introduction One major challenge in Urban Search and Rescue (USAR) is how to search a site for survivors without risking the lives of rescuers. In the past, robotic vehicles have been too large or heavy to use in collapsed or weakened structures. Furthermore, these vehicles require specially trained operators to remotely control them. As a result, USAR eorts such as at Oklahoma City are typically done manually. Two recent advances in mobile robot technology suggest that robotics may at last be practical for USAR. First, the hardware has matured. Lightweight inspection robots which can navigate through ducts and pipes are now commercially available for less than $20k. These systems still demand trained operators for teleoperation. However, progress in articial intelligence has reached the point where routine activities can be transferred to complete computer control, while the operator retains higher level control and supervision responsibilities. 1 Center for Robotics and Intelligent Systems, Colorado School of Mines, Golden, CO 80401
This semi-autonomous control paradigm simplies the demand on the operator, allowing a user to achieve the same results with less training, and to direct multiple robots at the same time. We are investigating methods for the semi-autonomous robotic examination of an area. This paper rst describes how animals search and forage. Insights from these studies leads to our formulation of three search strategies. Next, the paper presents how these strategies can be accomplished via reactive behaviors within a single semi-autonomous control architecture. The paper also reports on work in progress in implementing these search strategies on the Colorado School of Mines mobile robot, Clementine. The contributions of this paper are twofold. First, it oers a better understanding of search strategies. Second, it presents a novel organization of robotic search software with signicant advantages for USAR. This software architecture is expected to improve USAR search activities by automating systematic and opportunistic search thereby reducing operator fatigue and the need for specialized training. By reducing the need for continuous supervision, the operator may be able to concurrently direct multiple robots. Biological Foundations of Search Search and foraging skills in biological agents (humans and animals) have been investigated by behavioral and cognitive pyschologists; these studies provide insight into how a robotic search can be organized. Cross and Wellman (1985) demonstrated that children use two general search patterns for nding a well-dened target: comprehensive and selective. Comprehensive search is an exhaustive search, where the agent distributes the search eort evenly throughout the search space. In selective search, the agent infers the likelihood of the target being at particular locations and uses that inference to visit promising locations within the search space rst. Foraging behavior provides examples of methodologies for canvassing large areas under the constraints of time and energy. Unlike a true comprehensive search which covers the entire search space with even eort, foraging is a trade-o between win-stay and win-shift strategies according to Heth and Cornell (1985). A bee will forage heavily for nectar where it has encountered several owers in close proximity with nectar (win-stay), but if it nds only one or two owers with nectar out of many, it will ignore nearby owers and move a signicant distance away and resume foraging (win-shift). Essentially bees stay in an area as long as it is productive to do so. Systematic, Semantic, and Opportunistic Search for USAR The studies described above suggest that there are two basic categories of active search for USAR. The rst type partitions a large search space into subspaces and orders them by expected returns (likelihood of survivors, proximity, etc.).
LOCAL OPERATOR INTERFACE search pattern target alert teleop commands detection function sensor displays hardware configuration SEARCH Search Controller MANUAL robot CONTROL in place Navigation Scanning Examination REMOTE Figure 1: Layout of semi-autonomous, behavioral search architecture. We dene that type of search as semantic search, where the robot is directed to semantically similar areas of interest. The second is systematic search, a search eort which covers a search space, either comprehensively or by some sampling scheme. These two categories seem to correspond favorably to the natural division of cognitive abilities of the operator and remote robot. The operator is expected to perform semantic search because it requires abstract knowledge and reasoning abilities. The robot can execute the tedious, repetitive search for targets autonomously, alerting the operator when a potential target is detected. A combined systematic and semantic search initiative is not sucient for USAR applications. One advantage of human teleoperation is that the operator can sometimes notice targets even while performing another task, such as moving the robot to a new search location. Therefore a third category of search is needed to duplicate this function: opportunistic search. Under opportunistic search, the robot could autonomously run a reduced detection function on incoming data without distracting the operator or slowing navigational control. Semi-Autonomous, Behavioral Approach to Search Our approach uses a semi-autonomous control scheme as the framework for teleoperation. An overview of the architecture is shown in Figure 1. The operator can either manually control the robot or instruct it to perform previously programmed routine search activities without direct supervision. If the robot completes the activity (e.g., nds a potential target) or encounters an insurmountable diculty (e.g., cannot make progress due to a dead end), it contacts the operator and goes into a safe waiting state until the operator resumes control.
The operator can assume control of the robot at any time. The autonomous systematic and opportunistic searches are handled by a collection of reactive behaviors (Arkin, 1987 and Brooks, 1987) grouped into an abstract behavior (Murphy and Mali, 1995) called search. As shown in Figure 1, search has four components: a search controller which computes the win-stay or win-shift decisions based on the operator's input, a navigation behavior which has the robot maintain the desired search pattern, a scanning behavior which controls how the sensors mounted on the robot scan for targets and actually detects the targets, and an examination behavior which positions the robot to give the operator the best view once a target has been detected. It should be noted that the navigation and scanning behaviors can operate concurrently; i.e., the robot can turn its head back and forth while moving ahead if the sensing update rates permit it. The components of search execute independently of and concurrently with low-level behaviors that enable the robot to avoid obstacles, maintain speed, etc. The search controller is responsible for directing the navigation, scanning, and examination behaviors. It is the component that would encapsulate any mathematical models of search, such as those formulated by Cross and Wellman (1985), to compute when to shift or abandon the search, and adapt the detection function and/or resolution of the search pattern as needed. If the operator takes control of the robot to manually investigate a patch and then returns control to the robot, the search controller can remember the area that has been covered manually and adjust its pattern accordingly. The navigation behavior is responsible for computing the motor control commands (turn left, go forward) to maintain the search pattern. A special case of the navigation behavior is opportunistic search. Opportunistic search does not have a search pattern per se. In our implementation, it is invoked automatically when the operator teleoperates to next area of interest. The scanning behavior is responsible for controlling the sensor eector and executing the target detection function. This allows a robot to pan and/or tilt a sensor independently of the navigation behavior, making the software robot independent. The examination behavior suspends all navigation when the detection function discovers something. It alerts the operator, and attempts to move the robot and/or sensors to optimize the perceptual information about the detected target displayed to the operator. Work In Progress The search behavior has been implemented on one of the CSM mobile robots, Clementine, as a proof-of-concept. Clementine, shown in Figure 2, is a Denning MRV-4 mobile robot. She has a laser navigation beacon system for localization, a ring of 24 ultrasonic transducers for obstacle avoidance, and up to two color video camcorders xed in place on the robot for target detection. The navigation behavior currently supports one search pattern: sector-based.
Figure 2: Clementine in test search spaces. Others will be added later. In the sector-based search pattern, the operator positions the robot in the middle of the search space and instructs it to divide the area into slightly overlapping sectors based on sensor characteristics. The robot assumes autonomous control and turns to the rst sector. If it nds a potential target in the current sector, the search is suspended and the examination behavior is triggered. The examination behavior estimates the distance to the target, moves the robot, and waits for the operator to inspect the scene. If the operator directs the robot to resume the search, the robot looks in the next sector from its current position in order to save time and energy. If it does not detect new targets, the robot returns to the center of the search space and scans the next sector. When it has covered all sectors, it informs the operator that the search is now complete. The scanning behavior consists of a vision based detection function which looks for blobs in the image with one or more colors of interest. Figure 3a. shows an image taken from Clementine. The image in Figure 3b. shows the international orange and red (blood-colored) blobs that were detected in black. Discussion The implementation described above makes use of a simple search pattern and color-based detection function. Several additional issues must be addressed in order to transfer the concept into a reliable, useful software package for USAR. First, the selection of other behaviors to avoid obstacles, coordinate legged or tracked locomotion, etc., was ignored in this paper. These behaviors are well understood by the mobile robotics community and this is expected to be straightforward. Another issue is the choice of search patterns. A trained operator may be able to select a reasonably ecient search pattern given initial information, such as architectural drawings. However, the disaster can invalidate the expectations; for example, impassible rubble may eectively create two separate areas to be searched. The use of reactive behaviors allows the robot to continue searching
a. b. Figure 3: a.) View from robot camera. b.) Potential targets extracted by the detection function (in black). the reduced space but the search pattern for the anticipated large area may be totally inecient for the smaller area. We note that the inspection robot can provide information about its actual path and the search eciency to the operator and request a new search pattern be selected. Likewise, rubble may have reduced the utility of a particular robot conguration. As more intelligence is added to the system, it is conceivable that the computer at the local station could recommend new search patterns or hardware congurations. The cooperation between operator and robot outlined in this paper is limited to the division between them of cognitively challenging (semantic search) and routine (systematic and opportunistic search) activities. However, much more could be done to facilitate this cooperation. The software could be expanded to include a local expert system(s) to help an inexperienced operator quickly assess the situation and (1) select the most appropriate type of systematic search for a particular inspection robot and/or (2) best conguration of locomotion and sensors for a robot in that area. One such expert system, KNOBSAR, for advising USAR operators on which robots to deploy under what circumstances is currently under development at the Colorado School of Mines. Another aspect of cooperation is cooperation between multiple robots. If a robot encounters a high density of targets, additional robots could be transferred to that area to aid the search. This opens a new set of research issues about coordination and control, especially with a semi-autonomous paradigm. These issues include whether the robots' search eorts are better controlled by the operator or by a centralized articial intelligence at the local system, and how the robots will be programmed to act around each other (e.g., ignore each other or actively cooperate). The reader is directed to (Mataric, 1992) for a discussion of the levels of intelligence needed for cooperative multiple robots.
Acknowledgements This work is sponsored in part by NSF NSF Grant IRI-9320318, ARPA Grant AO#B460, the Colorado Space Grant Consortium, and an anonymous donor. The authors would like to thank Tyler Devore, Tim Flower, and Dale Hawkins for their help in designing and implementing the behaviors, and Maj. John Blitch for his helpful discussions about the USAR domain. References Arkin, R.C., \Motor Schema Based Navigation for a Mobile Robot: An Approach to Programming by Behavior", Proceedings of 1987 IEEE International Conference on Robotics and Automation, IEEE Computer Society Press, 1987, pp. 264-271. Brooks, R. A., \A Robust Layered Control System For A Mobile Robot", IEEE Journal of Robotics and Automation, IEEE Computer Society Press, vol. RA-1, no. 1, March, 1986, pp. 1-10. Cross, D.R., and Wellman, H.M., \Mathematical Models of Search," Children's Searching: The Development of Search Skill and Spatial Representation, H.M. Wellman, ed., Lawrence Erlbaum Associates, Hillsdale, NJ, 1985, pp.251-286. Heth, C.D., and Cornell, E.H., \A Comparative Description of Representation and Processing During Search," Children's Searching: The Development of Search Skill and Spatial Representation, H.M. Wellman, ed., Lawrence Erlbaum Associates, Hillsdale, NJ, 1985, pp.215-250. Mataric, M.J., \Minimizing Complexity in Controlling a Mobile Robot Population," proceedings of 1992 IEEE International Conference on Robotics and Automation, IEEE Computer Society Press, May, 1992. Murphy, R., and Mali, A., \Lessons Learned in Integrating Sensing into Autonomous Mobile Robot Architectures," submitted to Journal of Experimental and Theoretical Articial Intelligence. Somerville, S.C., and Haake,R.J., \The Logical Search Skills of Infants and Young Children," Children's Searching: The Development of Search Skill and Spatial Representation, H.M. Wellman, ed., Lawrence Erlbaum Associates, Hillsdale, NJ, 1985, pp. 73-104.