Canine as Robot in Directed Search

Alex Ferworn Alireza Sadeghian Kevin BarnumDept. of Computer Science Dept. of Computer Science Ontario Provincial Police

Ryerson University Ryerson University P.O Box 728,Toronto, Ontario, Canada, Toronto, Ontario, Canada, Bolton, Ontario, Canada,

M5B 2K3 M5B 2K3 L7E 5T5aferworgscs.ryersornca asadeghi(gscs.yerson.ca Kevin.Barnumn usmgovonca

Devin Ostrom Hossein Rahnama Isaac WoungangDept. of Mechanical Eng. Dept. of Computer Science Dept. of Computer Science

Ryerson University Ryerson University Ryerson UniversityToronto, Ontario, Canada, Toronto, Ontario, Canada, Toronto, Ontario, Canada,

M5B 2K3 M5B 2K3 M5B 2K3dostrom(gryersornca hosse n(~cs.ryersornca woungn(scfeson.ca

Abstract - One of the goals of rescue robotics has inevitably lead to the production of rubble often on abeen to allow machines to sense and act where vast scale. Speed, accuracy and safety are critical factorspeople cannot go. Deployed with various sensors and when searching for survivors in urban area turned tounder the guidance of trained operators these rubble.systems have shown great potential for findingvictims in rubble. However, the current state of Unfortunately search is a difficult and oftenmobility of all deployed rescue robots is such that dangerous process. While buildings collapse inthey can only be employed in very constrained predictable ways, the orientation, interaction and size ofcircumstances where surfaces are mostly regular and the rubble piles formed are often unpredictable. Spacesacceptable for various modes of tracked and wheeled or voids appear and disappear as the piles shift andmobility. We have been exploring an alternative settle. As time passes the rubble settles but theemploying the existing abilities of trained USAR probability of finding survivors is greatly reduced.canines to search rubble both quickly and effectively. Searching though rubble is time consuming becauseThis paper discusses one ofthe potential areas where footing is often unstable as are the structures that form ascanines might be augmented to deliver the desired a result of the initial collapse. There are usually nosensing and remote control ofa robot. obvious paths and the threat of secondary collapse is

sometimes a concern and often limits where searchersKireteywords:ar Canine Augmentch,USAtionR can go. Coordination of the search effort is made more

directed search,'rubble search, USAR. difficult by the rigors of simply moving around on the1 Introduction pile.

Rescue robotics is a relatively new field. It It is perhaps not surprising that first responders willprovides the hope that humans will not be put in turn to any system that can help with the process ofharm's way while searching for other humans who searching. Robotic technology has been applied toare trapped in wreckage. Numerous practical various common problems associated with the searchexamples of how robots can be used in urban search such as checking areas above the ground, dealing withand rescue (USAR) operations to assist in the search confined spaces and exploring areas that areprocess have been presented in the literature [1, 2, 3, contaminated. Their advantages are clear they can be4, 5, 6]. However, one goal that has remained elusive directly controlled by operators, they can be loaded withis the practical application of robotic technology to sensors and they do not place people's lives in danger.gain purchase on and move through real rubble.

A problem that continues to frustrate the applicationThe pervasive destruction caused by the Mexico of robots to rubble search is the same problem that slows

City earthquake (1985) and the destruction of the human searchers mobility. It has been the hope thatWorld Trade Center disaster (2001) have made it robots will become available that can be deployed andabundantly clear that the destruction of built-up areas

1-4244-1 160-2/07/$25.OO ©2007 IEEE.

quickly move around on rubble piles, reliably find audio streams transmitted from a trained USAR dog totrapped people and indicate their locations. The the search team employed in directed search. Thecurrent generation of rescue robots has not been able handler--via hand gestures and direct voice commands--to cope with real rubble often becoming stuck, provides commands to the dog. A simplified diagram ofbreaking, loosing mobility or otherwise failing. the service arrangement discussed in this paper is shown

in Fig. 1.One system that has proven its ability to

effectively move through rubble is the dogcomponent of US&R search teams. A dog's naturalagility and keen sense of smell allows dogs to moveover rubble very quickly. Unfortunately theeffectiveness of canine teams can be hampered by a Handlernumber of limitations associated with the mobility of Netmnd'the human on the team: v wIa Microphone* While a trained dog's agility can often let it R i mrmove quickly, a human handler must be more careful Min order to avoid injury. m* Dogs are not restricted from moving throughareas that a handler may not be allowed to enter for Caninereasons of safety. Thus the dog cannot search very far Figure 1. Trial Architecture.into these areas without losing contact with thehandler. The current generation of CAT mounts are based on* Dogs can often search and accurately clear human ski goggles and are commercially available fromcollapsed structures much more rapidly than a slow, dog-goes GmbH of Germany and sold under the brandinjury-prone human can follow. name "Pet-lenses" [7]. The goggles allow us to attach

appropriate technology to them but also provide theThe dog also limits the performance of the team benefit of being easily removed by the dog should the

because it can only "speak dog". Dogs indicate the goggles become snagged. An attached camera is shownpresence of human scent through a bark indication. in the figure below.Unfortunately, this does not communicate much moreinformation than the location of a casualty.

2 Canine Augmentation Technology

The Canine Augmentation Technology Project(CAT) was introduced in [9]. Our intent has been toprovide useful augmentation to canine teams insupport of US&R where direct interaction betweenthe canine and handler is precluded. Essentially, weAadd sensing and control services to the team. The Figure 2. CAT with Camera.CAT system is being developed to provide a platformfor services that are considered common in manyrescue robots but are generally unavailable for canine 3 Directed Searchsearch teams. In essence we wish the handler andother responders see what the dog sees, hear what the A relatively recent addition to Federal Emergencydog hears, know where the dog is and allow the Management Agency (FEMA) requirements for caninehandler to interact with and control the dog remotely. search teams listed in the Disaster Search Canine

Readiness Evaluation Process (DSCREP) [8] is the needThe project is a cooperative effort between the for "Direction and Control". In order to search areas that

Network-Centric Applied Research Team (N-CART) a canine handler may not access or to avoid hazardousof Ryerson University and the Provincial Emergency areas, a USAR canine must demonstrate the ability to beResponse Team (PERT) of the Ontario Provincial directed and controlled by the handler. The canine mustPolice (OPP). In this paper we will concentrate on the respond to audible and/or visual signals and move whereresults obtained in a CAT trial involving video and the handler wants it to go.

This is an interesting departure from common As the dog was directed between elevationspractice in canine search. Normally, search dogs recordings were made of both the view from the handlerwork in a self-determined, pattern that allows them to taken by an assistant collocated with the handler and thepick up a scent and follow it to likely areas view from the CAT system on Dare. A wireless receivercontaining trapped individuals. As the canine team was used to pickup the audio and video signals comingsearches, it moves in a complex interaction between from Dare. These were recorded on a common videothe dog and the handler, where the dog leads by scent camera. Dare was run through each of the elevationsand the handler provides correction and several times and was rewarded with play. The testingencouragement. The purpose of directed search is to took place on a bright June day at PERT headquarters inallow the handler to place the dog where there is a Bolton, Ontario, Canada.greater probability of picking up a scent.

SAR Dogs have not generally been trained tomove to specific locations with any precision by theirhandlers. As dogs are essentially autonomoussystems, they "decide" to move to a location through

canine must climb to required elevations at thedirection of the handler and wait for furtherdirections. In addition the dog must not climb on toelevations that have not been indicated. Theserequirements subtly change the interaction betweendogs and handlers. Rather than acting as a search Figure 4. Dare with CAT. Canine view inserted.team, the dog is being directed in the same manner asa rescue robot might be moved to a particularlocation. A typical DSCREP test environment is

Omade up of a series of elevations often fabricated 5 Observationsfrom wooden pallets or similar local material. Dogs For the most art, video and audio were remarkablyare typically directed to go right, left, move back and

c ccome~~~ ~fowad A tyia seu sso ntefgr clear considering that the dog was in motion. The voicebomelforw. rd. A typical setup is show in the figure of the handler and assistant were somewhat muffled but

distinct in the CAT recording and the breathing of thedog could be clearly heard and may provide a means ofmonitoring the status of canines employing such asystem in the future.

v Unlike the motion of many robots the dog was quitestable providing a very good base for mounting thecamera. Because the dog was able to move under its owncontrol between elevations, the handler could

Figure 3. Directed Search Setup. concentrate on planning the next move. This is animportant distinction between the performance of aCAT-equipped dog and a typical rescue robot. Because

4 Trl etals the dog is intelligent, is able to efficiently move across

rough terrain and is able to interact with the handler at a

In this trial we created a FEMA test environment high level, the canine team need not worry about thewith a series of appropriately positioned wooden minutia of control which continues to hamper the

pallets placed approximately 50m apart on loose performance of the current generation of rescue robots.

gravel. Over a series of months the SAR dog "Dare"was trained to respond to visual and voice commands The figures below provide some indication of what

the video feed is like. The image pairs present first thefrom a handler to comDly With the FEMA view from the assistant's camera and next the view fromreqireenfodreced each.Fo th traltheCA the CAT camera mounted on Dare. The views have beensystem was furnished with a wireless camera anda iesnhoie.Fgr n r iw ewemicrophone mounted between the goggle lenses with aeadhnlrwieDaei itn naple waitinga battery embeddled in the goggle strapping.

for the next command. Figure 7 and 8 are images Dare was being rewarded with play. A "Kong" toy istaken while Dare was in motion between pallets. used to reward Dare. In this interaction, the handler

threw the toy and Dare chased after it. The image shownbelow is a typical view of this event. The Kong remainsprominent in the field of vision of the camera and Dareis clearly tracking it visually while the Kong is in the airand Dare is running over the gravel at approximately 40KPH.

Figure 5. View of Dare on Pallet.

Figure 9. CAT camera tracking Kong.

...... ... ..... 6 Caveat EmptorMounting technology on a dog's head does not

mitigate all the problems associated with coming toterms with situational awareness through a narrow fieldof interaction. Many of the human-robot interaction

Figure 6. View from Dare on Pallet. issues concerning situational awareness at a disaster siteas presented through the interface of a rescue robot [9,10] are also present in a CAT human interface.Movement through a FEMA test course is not movementthrough a rubble pile. There are many physical andperceptual challenges that must be overcome.

7 Where a Dog LooksIt is likely that video and audio provided from a

canine located at the spot where it has detected aFigure 7. Dare Walking to Pallet. potential trapped human would be quite useful. As we

observed in this trial and various others we have run, thedog does not always look at scenes that rescuers mightfind useful. Indeed, it is not certain that video receivedfrom a moving dog on an active rubble pile would be asuseful given the erratic quality of the video presentedand its lack of context. Typically, when a canine ismoving quickly, following scent, the recorded video hasbeen very jittery and subject to sudden changes ofdirection and disorienting motion. From our observationsin this trial we know that dogs do not necessarily rely onvision when they are moving. This is fine for them but

Figure 8. View from Dare Walking to Pallet. would lead to problems for any observer attempting toDogs are not robots and as such it is important to follow the dog's progress in the pile.

keep them motivated with play. An unexpectedfeature was recorded from the CAT camera when

In addition to the directed search, we have run five became non-functional after Dare smashed it into a cartrials on three separate rubble piles involving door. Clearly more work is required to create aindependent search (the dog is free to search as it ruggedized version of CAT for deployment.wishes). Typically, video quality and reception varyimmensely as a dog moves through rubble, passes Referencesthrough brush, loose debris or climbs elevations. It isoften very difficult to determine exactly what thecamera is aimed at when a dog is moving. The signal [1] R. R. Murphy, "Trial by fire (rescue robots),"is occasionally lost and regained and the camera IEEE Robotics & Automation Magazine, vol. 11, 2004,orientation occasionally shifts as the dog either pp. 50-61."adjusts" the harness or simply chooses to movedifferently. [2] J. Casper and R. R. Murphy, Human-robot

interactions during the robot-assisted urban search andWhen a dog searches, its head is down as it is rescue response at the world trade center, IEEE

using scent to determine its direction of movement. Transactions on Systems, Man and Cybernetics, Part BWhen a casualty is found, the dog's head is raised as 33 (2003), no. 3, 367-385.it barks an indication to the handler. This isproblematic as the camera tends to either be pointed [3] R. R. Murphy, Marsupial and shape-shiftingat the ground or pointed at the sky more robots for urban search and rescue, IEEE Intelligentimportantly, it is not pointed at the casualty. Systems and Their Applications 15 (2000), no. 2, 14-19.

In directed search, the dog must look back at the [4] J. L. Burke and R. R. Murphy, Human-robothandler to receive a complete command as part of a interaction in usar technical search: Two heads arecommand is a gesture. When searching better than one, IEEE International Workshop on Robotindependently, the dog does not necessarily look at and Human Interactive Communication, 2004, p.App.the spot where a casualty may be located as sight is 307-312.probably not how the dog found the spot in the firstplace. This is problematic in that there may not be a [5] A. Davids, Urban search and rescue robots.direct translation between how the dog perceives the From tragedy to technology, IEEE Intelligent Systemscasualty and how that perception can be and Their Applications 17 (2002), no. 2, 81-83.communicated to a remote handler or search team.We believe that this problem can be partially [6] A. Ferworn, G. Hough, R. Manca, B.addressed through appropriate training but may be Antonishek, J. Scholtz and A. Jacoff, Expedients forbeyond the capabilities of some dogs. marsupial operations of usar robots, IEEE International

Workshop on Safety, Security and Rescue Robotics

8 Conclusions (SSRR06), 2006, p.App.We have demonstrated that canine video is both [7] http://www.pet-1enses.com.

stable and potentially useful if certain factors can beovercome. We are working to determine what factors [8] FEMA, "Disaster search canine readinesscan be overcome through the application of evaluation process," Federal Response Plan Emergencyappropriate technology and what factors require Support Function 9, Federal Emergency Managementadditional training. Unlike rescue robots, the dogs Agency, 2006, p. 45.employed in the system are intelligent and take careof their own movement. This has the potential [9] J. Drury, L. Rick, A. Christansen, Z. Byler-advantage of allowing more searchers to concentrate Walker, A. Maggi and D. R. S. Smith, Evaluatingon searching rather than dealing with the mundane human-robot interaction in a search-and-rescue context,but difficult task of low-level control. Rubble is a Performance Metrics for Intelligent Systems Workshop,very demanding environment for search teams and 2003, p.App.their equipment. This is no different for CATequipped canine teams. During this trial several [10] H. A. Yanco and J. Drury, "Where am i?b"systems became non-functional. One camera was lost Acquiring situation awareness using remote robotas it was ripped from the goggles it was mounted on. platform, IEEE International Conference on Systems,Perhaps tellingly this loss occurred in the relatively Man and Cybernetics, 2004, p.App. 2835-2840.benign environment of tall grass. Another camera