Figure 1: The micro-rover Bujold is deployed from inside the car-like Silver Bullet through gate in rear, much like a kangaroo carrying its young. (Gr

Transcription

1 Marsupial-like Mobile Robot Societies Robin R. Murphy, Michelle Ausmus, Magda Bugajska, Tanya Ellis Tonia Johnson, Nia Kelley, Jodi Kiefer, Lisa Pollock Computer Science and Engineering 4202 East Fowler Avenue, ENB 118 University of South Florida Tampa, FL Abstract Recent work in mobile robotics has produced a novel multi-agent society, dubbed in this paper as marsupial robots. In applications such as Urban Search and Rescue, humans are often involved in control and communicating with the robot teams. As a result, marsupial robots oer rich, complex opportunities for exploring issues of representation; adaptive control, communication, and cooperation strategies; and for reasoning about the roles of agents for a task and environment. This paper denes the marsupial robot team concept and proposes a taxonomy of roles for each agent. It describes the team members (Human, Dispensing Agent, and Passenger Agents) in terms of the characteristics of more traditional agent societies. The marsupial concept is compared with related work in multi-agent robot teams. The paper also discusses the results of an investigation where environmental and task constraints dictate the deployment of Passenger Agents, and the resulting heuristics. 1 Introduction Recent work in mobile robots [1, 3] has generated a novel type of heterogeneous agent team: marsupial robots. The term marsupial robot connotes a 1

2 Figure 1: The micro-rover Bujold is deployed from inside the car-like Silver Bullet through gate in rear, much like a kangaroo carrying its young. (Grid lines are at 2ft intervals.) larger \mother" robot carrying one or more smaller \baby" robots much as a kangaroo mother carries her young. However, the relationship between the mother robot and the babies can be much more complex than just the mother providing transportation. Following the biological analogy, the mother robot can provide power (\food") or help (rescue the baby, communicate suggestions or warnings). One such multi-robot organization is illustrated in Figure 1 by the Silver Bullet and Bujold team demonstrated at the 1997 AAAI Mobile Robot Exhibition, and featured in [7]. The need for marsupial robots naturally arises in domains where microrobots are used to explore remote locations. Two terrestrial examples of these domains are Urban Search and Rescue (USAR) and underwater robotics. A micro-rover generally cannot transport itself to the remote location without a serious penalty in time or power. Therefore, it is desirable to have a larger robot transport the micro-rover and perhaps serving as a base station, containing power recharging abilities and an intermediate communication link. The larger robot can also continue an active role in the team by providing additional viewpoints of what the micro-rovers are doing or their progress. In both USAR and underwater applications, maintaining a link to humans is important. A human supplies the decision making capabilities which have not yet been successfully duplicated. But, a human is important for 2

3 other reasons. In the case of USAR, it is desirable that once a robot has found a survivor, it permit the survivor and human rescue worker to talk. The worker needs to perform triage on the survivor's condition, and should collect information as to the possible number and location of other survivors, presence of hazards (gas leaks), etc. While the human agent is required for the mission, it may be undesirable to have the human serve as a teleoperator. Teleoperation is notoriously dicult, and requires a signicant amount of training. It is impractical to attempt to train USAR workers on robots that may or may not be transported to them for an emergency. Plus, the unfamiliarity compounded by the stress of the situation may cause the operator to make mistakes. Instead, advances in semi-autonomous control of mobile robots can mitigate the cognitive fatiguing aspects and permit the human to concentrate on other, non-automated tasks. This paper addresses the implications of human-marsupial robot teams for agent societies and reports on preliminary research in dening roles and control methods. It denes a marsupial robot team as a collection of mobile robots, where one or more robots are at least temporarily physically dependent on another for directives, transport, power, communication, etc. In order to distinguish a marsupial robot team from the biological analogy, the \mother" agents will be referred to as Dispensing Agents. Although they play other roles than simply transportation, it is the one role that this type of agent must be able to play. Likewise, the \baby" agents will be called Passenger Agents. Our research with marsupial robot teams is concentrating on 1) how to represent the abilities of agents within a heterogeneous team, 2) what are the roles of the Dispensing Agent and how it can change roles as needed, and 3) how the Dispensing Agent can reason about the capabilities of the Passenger Agents for a particular task and environment. This paper makes several contributions towards our research agenda. First, it denes a novel mobile agent society and the roles of the three different types: Human, Dispensing Agent, and Passenger Agent. It discusses these roles in terms of the characteristics of more traditional agent societies. Second, it reports the results of an empirical investigation of how the Dispensing Agent can autonomously determine when to deploy the Passenger Agents for a task. This provides both a practical set of heuristics which can be used by any marsupial team and a foundation for continuing eorts in our research. In addition, the paper provides an example of an implementation 3

4 of a marsupial team for the USAR domain. The work in marsupial agents is compared with related work in multi-agent robot teams, and directions for continuing work are derived. 2 Roles in Human-Marsupial Teams One distinguishing feature of Human-Marsupial teams is that the arrangement does not reect a xed hierarchy of intelligence, i.e. the Human Agent does not control the Dispensing Agent which does not control the Passenger Agent. Instead, the three types of agents can shift roles and control topology with circumstances. This section discusses the set of roles of the types of agents in terms of control and communication topology, cooperation strategy, and choice of goals following the taxonomies of [2, 5]. 2.1 Marsupial Team A marsupial team consists of a Dispensing Agent and one or more Passenger Agents. The physical marsupial-like arrangement does not limit the programming of the Passenger Agents. The Passenger Agents can function with any reasonable combination of multi-agent control, cooperation, and goal strategy. For example, once deployed, the Passenger Agents can function under distributed control, with individual goals (e.g., dierent regions to explore), and non-active cooperation [5]. Or they could be centrally controlled by the Dispensing Agent or the human. The Passenger Agents may also be heterogeneous themselves. What makes the marsupial team unique agent is the presence of a Dispensing Agent with dierent viewpoints, sensors, and cognitive abilities from the Passenger Agents. The Dispensing Agent can assume one of at least four distinct roles: 1. Coach. In this role, the Dispensing Agent deploys semi-autonomous Passenger Agents, then actively attempts to aid the Passenger Agents. This is equivalent to a coach running up and down the sidelines shouting advice (\watch out for x"). For example, in a USAR situation, the Dispensing Agent would take advantage of its dierent viewpoint and sensors to help guide the Passenger Agents. The dierent viewpoint and sensor suite might be able to evaluate terrain preferences and determine goals. Also, the Dispensing Agent can attempt to maximize its utility 4

5 to the deployed Passenger Agents by moving itself and sensor eectors to maintain the most useful viewpoints. Note that this information can be at very high levels of abstraction, with token-like representations to reduce communication bandwidth. The Passenger Agents are semi-autonomous and attempt to take advantage of the suggestions or broader perception, but still have complete tactical control of themselves. 2. Manager. The Coach role assumes that the Passenger Agents are working under a distributed control scheme, and that they integrate any information from the coach individually. The Manager role assumes a centralized control scheme, where the Dispensing Agent explicitly directs each Passenger and coordinates those actions with that of the entire team. See Figure Messenger. The Dispensing Agent can remain tightly coupled to the Passenger Agents, either passively or actively, by acting as a relay station for communication. The dierence between the Messenger and Coach role is the seat of autonomy. In the Coach role, the Dispensing Agent is making the recommendations. In the Messenger role, the Dispensing Agent does not communicate with the Passengers, it merely tries to facilitate radio contact to the human operator or for centralized communication by the Passengers. 4. Courier. Although the marsupial team in Figures 1 and 2 shows the Passenger Agent, Bujold, physically tethered to the Dispensing Agent, Silver Bullet, Passenger Agents do not have to be tethered. In that case, the Dispensing Agent can serve to deliver sets of Passenger Agents to dierent locations, then continue with tasks unrelated to the Passengers, and eventually come back and pick up the Passengers. In this role, the Passengers are only loosely coupled with the Dispensing Agent. It is expected that the roles of the Dispensing Agent will change throughout the phases of a task. The Dispensing Agent might deploy the Passengers near a pile of rubble. The rubble may be such that the Passengers cannot see a void to explore or avoid box canyons. The Dispensing Agent could either 5

6 Figure 2: Silver Bullet in the Manager role, directing Bujold through an obstacle course. coach or manage them over rough terrain to voids to enter and explore. Once they reach their void, the Passenger Agents may become out of sight. Then the Dispensing Agent would act as a Support Group, receiving and relying signals, providing data fusion for centralized control, etc. 2.2 Human Interaction In many applications, especially USAR, it is useful to have a human interact with the marsupial team. First, human input might be needed to dene the overall mission in terms of what roles and activities the robots were to perform, and the criteria for termination. This is referred to as taking a Mission Controller role. Second, the output from the Passengers via the Dispensing Agent or other Passenger Agents may need to interact with the human operators. For example, the human may notice cues indicating a survivor that were missed by the robot. More likely, the robotic agent would request a human conrm an ambiguous cue of a survivor and assume a shared control mode. It is important to note that the human can interact with either the Passenger Agents, the Dispensing Agent, or both. This is the Task Guidance role. 6

7 3 Silver Bullet and Bujold This section describes a marsupial team developed explicitly for USAR: Silver Bullet and Bujold. The robot team was used for the experiments described in the next section. Silver Bullet, the Dispensing Agent, was designed and built by students as part of an NSF-sponsored Research Experience for Undergraduates. Silver Bullet's chassis is a Fisher Price Power Wheels children's jeep equipped with an on-board computer and radio ethernet link, six sonars for navigation, and a panning camera plus thermal sensor for detection of survivors. The onboard computer is a Pentium class AMD 133MHZ processor with 16MB of RAM, and 700MB hard drive. She was originally operated under MS-DOS, but now uses Linux. Power is supplied by either household current or one lightweight 12V motorcycle battery. Silver Bullet is sturdy enough to carry up to 70 more pounds of payload. Silver Bullet has numerous algorithms for navigation and search. Autonomous control of both Silver Bullet and Bujold is behavior based. Bujold, the Passenger Agent, is a commericially available tracked chemical inspection robot built by Inuktun Services of Canada. Bujold has a camera and two headlights. She has no on-board processors and is controlled via a teleoperation interface. It can physically change its height to \sit up" or \lay at." The camera can tilt independently of the physical conguration. A microphone was added for hearing survivors, along with a video transmitter to permit images to be sent to a human without the time delay introduced by going through Silver Bullet. Silver Bullet has a space in the chassis large enough to accommodate Bujold in a at position. Bujold stays in the compartment until a plexi-glass gate is lowered, and she is commanded to move forward. Bujold is physically connected to Silver Bullet at all times through a 100 foot long power and data tether. The tether is on a self-feeding spool. Silver Bullet carries two 12V batteries to serve as a dedicated power supply for Bujold. Also Silver Bullet has a second framegrabber to process images from Bujold independently of its own camera. Either a human operator of Silver Bullet can completely control Bujold through the teleoperation interface. A human can connect to Silver Bullet through radio ethernet and interact with both robots simultaneously. 7

8 4 Autonomous Deployment One of the research objectives is to identify and automate the transitions between roles in the Dispensing Agent. Ideally, the Dispensing Agent would be tasked to autonomously carry a set of Passenger Agents (Courier role) to a location, deploy the agents, and then serve in one of the other three roles (Coach, Manager, Messenger). Initial work has focused on developing a strategy for when the Dispensing Agent should deploy the Passengers. An investigation was conducted with the Silver Bullet-Bujold team in seven representative scenarios. In each scenario, the robots had to move to within 1 meter of a location where a victim (a mannequin) was expected to be. In practice, a USAR team may not have such expectations, but a specic target location was specied to make the tests repeatable. The scenarios were conducted in the Colorado School of Mines Mobile Robotics/ Machine Perception Laboratory (26 x 24 feet). Data was collected for each scenario for 1) Silver Bullet navigating alone, 2) Bujold navigating alone, and 3) Silver Bullet and Bujold acting as a team. The robots were teleoperated in order to ensure repeatability of path. Five data sets for each robot conguration for each scenario was collected (105 data points). The data compared the time it took the robots individually to navigate via teleoperation with the time for the marsupial team. The seven scenarios covered spatial congurations favorable for navigation by a large robot and for a micro-rover. Scenario 1 was an open space, which favored the larger, faster Silver Bullet acting alone. Silver Bullet outperformed Bujold by a margin of 2:1. Since there would be no reason to deploy the Passenger until Silver Bullet arrived, no data was collected on the marsupial team arrangement. Scenario 2 simulated rough terrain (rocks, rubble, loose wood and metal rods). In three runs, Silver Bullet got stuck on the rubble, while in another 3 runs, Bujold got stuck. Eight data sets were collected instead of ve to compensate for the aborted runs. The results showed that when Bujold did not become stuck, it was able to reach the target location twice as fast as Silver Bullet operating alone and three times as fast as the marsupial team conguration. Scenario 3 provided an area which was navigable by the robots, but would require the larger robot to have to slow down and make many turns in order to avoid obstacles. In this scenario, Bujold again did better that than the other two congurations, but 8

9 Figure 3: Scenario 4: Silver Bullet and Bujold have to take dierent routes if acting independently. the average marsupial and Silver Bullet times were tied. Scenario 4 provided a comparison of routes. Obstacles were arranged such that Bujold could go through a narrow opening and make a straight path for the target location, while Silver Bullet had to veer and go around a long wall of obstacles before resuming direct progress. Bujold acting alone and the marsupial conguration reached the target location with an average time dierence of 8 seconds. Silver Bullet was 50% slower. Scenario 5 showed similar results, only in this case, the scenario combined obstacles and rough terrain. Scenarios 6 and 7 created situations where Silver Bullet could not reach the target location by any path, and only Bujold and the marsupial conguration could. Bujold acting alone tended to do about 20 Overall, the data showed that for distances over the length of the tether, 9

10 Figure 4: Scenario 6: A dicult course that can't be done by Silver Bullet alone. Silver Bullet is prevented from going beyond the simulated ledge. 10

11 Bujold acting alone could almost always reach a survivor faster than Silver Bullet (1:52 vs. 2:45 minutes) or Silver Bullet and Bujold acting together (avg. 2:08 minutes) This is reasonable given the agility of the smaller robot. However, it was also noted that while Bujold was faster, the further the distance covered the more signicant the drain in battery power. The dierence in time between Bujold operating alone and acting in a marsupial team was fairly small, once the delay in physically deploying Bujold is eliminated. This suggests that the deployment strategy should include the power costs versus time gains in determining the appropriate point to deploy the Passenger Agents. Otherwise, the Passenger Agent could arrive at the target location with no energy to conduct the search and rescue activities. While the time data is informative, other lessons learned from the experiments were useful. In Scenarios 6 (Ledge) and 7 (Impasse), Silver Bullet was unable to reach the expected location of a survivor due to simulated obstacles and drop os. Without Bujold working with Silver Bullet, the mission could not have been accomplished without draining all the power for Bujold. It should also be noted that Scenario 2 illustrated a potential advantage of having a robot with a wide eld and range of view: the robot can assess the terrain and determine whether it is safe to deploy the Passengers or not. Based on the data collected, the following heuristics for the deployment of a Passenger Agent(s) have emerged: If the Dispensing Agent cannot make any further progress (is blocked), and the goal region to investigate is within tether (or communication) distance constraints, deploy the Passengers. If the Dispensing Agent cannot make direct progress (has to veer away to nd an alternate path), but the Passenger Agent can, and the Passenger is within tether constraints, deploy the Passengers. If the Dispensing Agent is within the tether constraints of the goal region and the ratio of Dispensing Agent travel time to Passenger Agent power consumption is high, then deploy the agents. Note that if the Dispensing Agent is not blocked, it can deploy the Passengers and while they make rapid progress to the goal region, it can catch up to them. At this time, we are considering encapsulating these heuristics with fuzzy rules. The qualitative nature of the heuristics suggests that a satisfactory utility function would be dicult to construct. 11

12 5 Related Work Robots which can physically transport other robots were described earlier in [1], but do not appear to have been completely implemented. In that application, a large Cybermotion robot, called MACS, was intended to carry one or more small robots (IS Robotics' R3 platforms named RACS) to various areas which needed to be surveyed for radiation. Once dropped o, the RACS robots were to conduct an autonomous survey of a room and then return to the start point to wait for pickup. The MACS would also drop o radio communications beacons to transmit the results of the surveys in progress to the MACS. As such, this application can be described as a marsupial team with the Dispensing Agent serving in the Courier role. The Passenger Agent equivalents, the RACS, used a distributed control scheme, where each agent had its own goals. Communication was one way and limited to data collection rather than control information, with the MACS acting as the central collection point. Also, the eort included a human agent; it was intended to have a human operator use a GUI to specify the areas to be investigated by the MACS/RACS team. The writeup [1] concentrated on the implementation details rather than on the overall contribution to agency theory. The use of marsupial robots, especially the control of the Passenger Agents, has similarities with swarm robots. However, swarm robots are generally homogeneous in both hardware and software capabilities, and do not accommodate human interaction. Likewise the multi-agent teams developed for RoboCup and other game playing robots have only limited relevance to marsupial robots. Game playing agents have dierent roles to play, and may be heterogenous both in roles (goalies, defender, striker, etc.) and in physical resources (sensors). However, they are not physically dependent on each other, do not have broad range of hybrid control regimes (competition rules generally restrict control to be either purely centralized or distributed), and do not permit human interaction. Swarm and game-playing agents provide methodologies to be used by the marsupial team. Other work of interest is [3, 4, 6]. Those eorts focus on the development of an intelligent assistant agent to aid a human interacting with multiple robots. The intelligent assistant would be even more helpful for humans working with marsupial robot teams due to the increased complexity and magnitude of roles and control schemes. 12

13 6 Conclusions and Open Issues In summary, a marsupial team consists of two type of agents: at least one Dispensing Agent and one or more Passenger Agents. In addition, a particular domain such as Urban Search and Rescue may require the involvement of a human. This leads to a complex set of roles for agents in the society. The Dispensing Agent must meet the role of being a Courier, but it may also serve as a Coach, Manager, and Messenger. The Human may be a Mission Controller or a Task Consultant. Furthermore, these roles change depending on the phase of the mission. One such marsupial team (Silver Bullet and Bujold) has been built and demonstrated for USAR applications. Preliminary work with the team in agency concentrated on dening heuristics to enable Silver Bullet to autonomously determine when to dispense Bujold. These heuristics suggest that the task and environmental constraints will have to be known or projected in order to make a reasonable decision. Open issues include when each agent should change other roles, and how to represent these roles and information to be communicated between agents. Current work is concentrating on investigating more aspects of marsupial teams for the USAR domain. Recent work has addressed the need for intelligent assistance in order to manage the information overload provided by the multiple members of a marsupial teams [3]. One issue specic to USAR are the deployment of more micro-robots, o tether which can be left behind, go to a survivor, and be a \personal robot." The robot would maintain audio communication between rescuers and the survivor, transmit bio data, possibly carry medicine, etc. Another issue being considered is the survivor as a member of human-marsupial team. In this context, the survivor may not be adding any cognitive abilities. Instead, the impact would be to inuence the behavior of the robots, since a survivor could be scared of robots to begin with, disoriented and in shock, and react negatively to the attentions of a rescue micro-rover. Acknowledgments This work is supported in part by NSF grants CDA and NSF REU Site Grant EIA The authors would like to thank Julian Martinez, 13

14 Dr. Karen Tichenor, and Aaron Gage for their help in conducting the experiments, Dr. Harry Hedges and LTC John Blitch for their feedback and encouragement, and the previous REU team who constructed Silver Bullet. References [1] Anderson, M.O., McKay, M.D., Richardson, B.S., \Multirobot Automated Indoor Floor Characterization Team," Proceeding of the 1996 IEEE International Conference on Robotics and Automation, Minneapolis, MN, April, 1996, pp. 1750{1753. [2] Dudek, G., Jekns, J., Milios, E., and Wilkes, D., \A Taxonomy for Swarm Robots," Proceedings of the 1993 IEEE/RSJ International Conference on Intelligent Robots and Systems, Yokohama, Japan, July, 1993, pp [3] Murphy, R., and Rogers, E., \Intelligent Assistance for Human Supervisory Control of Multiple Robots," SPIE International Symposium on Aerospace: Robotic and Semi-Robotic Ground Vehicle Technology, April 13-17, Orlando, FL, [4] Murphy, R. R., and Rogers, E., \Cooperative Assistance for Remote Robot Supervision," Presence, special issue on Starkfest, vol. 5, no. 2, Spring 1996, pp [5] Premvuti, S., and Yuta, S., \Considerations on the Cooperation of Multiple Autonomous Robots," Proceedings of IROS 1990, pp [6] Rogers, E., Murphy, R.R., and Ericson, B., \Agent-Based Expert Assistance for Visual Problem Solving," Autonomous Agents 97, Marina del Rey, CA, Feb. 5-8, 1997, pp. 156{163. [7] \Silver Bullet and Bujold go where humans fear to tread," Robot Science and Technology Magazine, premier issue, February,