Submitted to the International Conference on Intelligent Robotics and Systems (IROS) Las Vegas, October, 2003

Subsurface Surveying by a Rover Equipped with Ground-Penetrating Radar Timothy D Barfoot1 , Gabriele M T D’Eleuterio1 , and Peter Annan2 1

University of Toronto Institute for Aerospace Studies, 4925 Dufferin St., Toronto, ON, Canada, M3H 5T6 2 Sensors & Software Inc., 1040 Stacey Court, Mississauga, ON, Canada L4W 2X8 [email protected], [email protected], [email protected]

Abstract We discuss our experiences in integrating a commercial off-the-shelf ground-penetrating radar unit with an all-terrain rover. Straight-line subsurface surveys were generated in a fully autonomous manner using odometry and a simple visual servoing technique. Survey results for various terrains are presented. We discuss the configuration of the integrated system and make recommendations for both Martian and terrestrial applications.

1

Introduction

Ground-penetrating radar (GPR) is a popular tool that has been used for utility mapping, structure assessment, archaeology, geological characterization, and military applications such as land-mine detection [4]. By transmitting an impulse and measuring the time of flight until a reflected pulse is received, the depth profile of the various subsurface layers/objects may be inferred. A map of the subsurface can be built up by combining depth profiles from a large number of points on the surface. It can be time consuming and tedious for a human operator to take the thousands of GPR measurements needed to build up subsurface surveys. Moreover, for extraterrestrial applications, this process will need to be automated. To this end, we equipped a 15 kg rover with the 1000 MHz version of the OEM Noggin Plus GPR manufactured by Sensors & Software Inc. (Mississauga, Ontario). Our goal is to fully automate the process of building threedimensional subsurface maps. To date, we are able to construct two dimensional maps autonomously, where the rover drives along a path and takes GPR measurements at predefined intervals. We are particularly motivated by the possibility of using GPR on a Mars rover. Scientists are very keen to investigate the stratigraphy of Mars, particularly the polar layered deposits, in order to determine its

Figure 1: Rover conducting subsurface survey using ground-penetrating radar (in tow). climate history [10]. The straight-line surveys we will present here can be accomplished with a level of autonomy comparable to that of the Sojourner rover on the Mars Pathfinder mission [9]. The potential to yield a large amount of scientific data at a relatively low level of autonomy suggests GPR is a good rover-based science package for future missions. Rover-based GPR would also complement an orbiter equipped with synthetic aperture radar [7] and could provide useful information in the selection of sites to drill [14] or burrow [15] into the Martian surface. This paper is organized as follows. We first discuss related work followed by a detailed description of our system configuration. We then discuss the algorithms used to carry out straight-line surveys with results from experiments. Integration issues are discussed and recommendations made for future roverbased GPR work.

2

Related Work

Although there is a large body of literature on mobile robotics, our review here will focus on research that

seeks to integrate scientific instruments with rovers for planetary exploration. Beck and Osborn [3] describe similar goals to ours, the automation of GPR surveying, but concentrate on modelling of sensor data rather than integration with a rover. Arcone et al.[1] used a higher-frequency radar mounted on a large tracked vehicle to measure Arctic lake ice thickness, claiming some advantages over typical GPR frequencies for this application. Foessel et al.[5] describe Antarctic trials with a sled-mounted GPR to search for meteorites at the ice-snow interface. They used a 500 MHz unit but recommended a higher frequency to increase resolution for their application. Bapna et al.[2] discusses the first long-range semiautonomous operation of a science rover in the Atacama desert in Chile. The Nomad rover looked for meteorites using high-resolution stereo cameras and an eddy-current sensor. A meteorite was later found autonomously by Nomad during trials in Antarctica [12]. Pedersen [11] discusses the level of autonomy needed to place rover-based science instruments precisely against rocks for future Mars missions. We feel this work is very important for other science instruments but that GPR may offer good science with less sophisticated autonomy. Grant et al.[7, 8] describe a specific GPR configuration under development for rover-based Mars missions. They discuss both paired dipole antennas and a “rat-tail” type antenna that drags behind a rover. Foessel et al.[6] describe a radar instrument which can be used to map the subsurface and safeguard a rover by detecting such hazards as crevasses.

3

Hardware Configuration

Our system consisted of a 15 kg Argo Class Rover designed and constructed at the University of Toronto Institute for Aerospace Studies and a commercial OEM Noggin Plus GPR unit built by Sensors & Software Inc (see Figure 1). The rover is based on a Tamiya TXT-1 truck kit which was heavily modified. The main features include a Pentium III 700 MHz computer running a custom distribution of Linux, wireless ethernet communications operating at 2.4 GHz, encoders on all four wheels with a resolution of 512 ticks per revolution, steerable colour cameras with 640 × 480 resolution, sonar range finders on the front and back to detect obstacles, a two-axis inclinometer, and a chassis capable of independently controlling the payload platform to have it remain level on uneven terrain. The rover is powered by 34 Ah of NiMH batteries stored in the wheels to keep the center of mass low to the ground. Software access to the rover sensors

and actuators is provided by means of TCP/IP services which may be called by programs running onboard the rover or off-board on another computer. Typically, higher-level programs run off-board and time-critical programs run on the rover. The rover is approximately 50 cm long, 35 cm wide, and 45 cm high. The OEM Noggin Plus GPR unit is approximately 2.3 kg, 30 cm long, 15 cm wide, and 11 cm tall. Its center frequency is 1000 MHz and is capable of imaging several meters into the ground. The unit draws 0.7 A at 12 V and communicates with the rover over an RS232 connection at 115 Kb/s. A bottommounted skid plate allows the unit to be dragged along the ground, thus minimizing the air-ground interface. The GPR was mounted behind the rover on extendable tubular cantilevers. A parallel-arm structure was employed to keep the unit flat on rough terrain. The GPR was mounted on the rover using all plastic hardware to minimize unwanted reflections. The main DC drive motor for the rover was located at the front, as far from the GPR as possible.

4

Software Configuration

Given the possibility of using GPR for Mars missions, we were motivated to develop techniques wherein the rover could accomplish the task without the aid of external sensors such as GPS. It turns out that for straight-line surveys a good sensor combination is to use wheel encoders to perform odometry and a single camera to perform visual servoing towards (or away from) a stationary landmark. The scenario we had in mind was a rover driving in a straight line away from a visually identifiable lander. Should a rover need to travel beyond line-of-sight range with its lander, a sun sensor or gyro could be used to provide orientation information. The primary test area we used was a large gravel surface on which the rover did experience slippage fairly often, as well as considerable vibration when the rover was moving. This necessitated that the rover stop to collect camera images to avoid blurring (low light levels required long exposure times). Distance travelled was computed by integrating the average of the four wheel speeds using a second order backwards integration scheme at 10 Hz. Given the resolution of the encoders and wheel diameter, distance travelled could be measured to millimeter precision (with no slip). Discussions of slippage issues will be left to a later section. The GPR records were triggered based on distance travelled. This was done by evaluating whether the odometry exceeded a prespecified distance interval.

60 40 20 0 −20 −40 −60 −80 20

40

60

80

100

120

140

160

Raw GPR Signal Strength

Raw GPR Signal Strength

Raw GPR Signal Strength

60

600

80

−100 0

Low Freq Noise from Servomotors

Spikes due to Wireless Ethernet Radio

Baseline Background Noise 100

400 200 0 −200 −400 −600 0

20

40

Time [ns]

60

80

100

120

140

160

40 20 0 −20 −40 −60 0

20

40

60

80

100

120

140

160

Time [ns]

Time [ns]

Figure 2: (left) Background noise on GPR receiver (100 trials of 160 ns overlayed). (middle) Spikes due to wireless ethernet communications (100 trials of 160 ns overlayed). (right) Typical low frequency noise due to steering servomotors (1 trial of 160 ns). If it was exceeded, a GPR record was triggered. This was done in a polling mode at a frequency of 10 Hz. We attempted using intervals of 10 cm and 2 cm. Each GPR record consisted of 201 sample points spanning a 20 ns window. This would allow reflections to a depth of a few meters to be recorded. Driving the rover in a straight line on loose terrain was accomplished using a simple visual servoing technique. To date the rover segments its camera image based on colour. We placed a large blue box approximately 35 m from the rover’s initial position. The colours of the raw RGB image were histogrammed into 30 bins using the rgb2ind routine in Matlab 6.0. We identified a pixel as being blue if ρ2 + γ 2 + (β − 1)2 ≥ τ

(1)

where (ρ, γ, β) are the normalized RGB values and τ = 1 is a threshold value. The Matlab roicolor routine was used to create a binary image where the “on” pixels represented blue and the “off” pixels were not. The geometric centroid of the “on” pixels was computed which served as the rover’s target coordinates. It tried to keep this target in the center of its image (left-right only) throughout the experiment. This colour segmentation technique will later be replaced with a more sophisticated vision system but it sufficed as a proof of concept. The process of collecting a GPR survey consisted of a loop in which the rover captured an image and computed the location of the blue target, planned a trajectory that would bring the target towards the image center, and then executed the trajectory. A trajectory consisted of driving along a path of curvature, K, at speed, v = 0.1 m/s, for time, T = 10 s. A nonlinear controller [13] was used to execute the trajectory and the state estimate (position/orientation) was computed by odometry as described above. The curvature was given by K = k(xcm − xo )

(2)

where k = −0.05 was a gain, xcm was the left-right location of the blue target, and xo = 120 was the left-right center of the image (which was 240 pixels wide by 320 pixels high). This sense-plan-act loop was repeated a fixed number of times. Typical trials involved 35 repetitions making the GPR surveys 35 m long.

5

Experiments

The GPR was operated in a passive listen-only mode to determine if subsystems on the rover would cause noise on the GPR receiver. Figure 2 (left) shows the background noise we recorded on the receiver with no radio communications and all motors off. The envelope of the background noise had signal magnitude 120 (of a possible 216 = 65536 on the analogto-digital converter). Figure 2 (middle) shows spikes that occur when the 2.4 GHz radio communications were operating, which increased the noise envelope by an order of magnitude to 1200. Figure 2 (right) shows a low frequency oscillation that was recorded as a result of the rear steering servomotor being turned all the way to one side. The servomotor noise did not increase the noise envelope significantly. We tried moving the GPR unit forwards/backwards with respect to the rover body but did not find a significant difference in the recorded noise with radio communications on and motors running. The DC main drive motor did not have a measurable affect on the background noise. Overall the radio communications had the largest effect on the GPR receiver but even this was tolerable. We conducted several trials of the system in a large 50 m diameter domed structure, MarsDome, at the University of Toronto. This facility is currently under development but will eventually be a large-scale robotics test facility for planetary rover experiments. Currently, the 40 m diameter test area consists of a 15 cm thick rough concrete foundation covered by 15 cm of loose gravel with grainsize of 2 cm.

Subsurface Survey on Gravel/Concrete

x 10

Subsurface Survey on Gravel/Concrete

4

2

3

3

4

5

0

6

1

Time [ns]

4

8 10

Bottom of Concrete Sublayer

0

12 −1

−1

14

Wheel slippage

8

2

6

1

7

Gravel Surface

2

15 cm alum angle

2

Time [ns]

Concrete Surface

2 cm steel pipe

1

4

x 10

3

16

−2

−2

18

9 10

20

−3

10 20 30 40 50 Distance Along Ground [m]

Signal

−3

6

12 18 24 30 Distance Along Ground [m]

Signal

Figure 3: Surveys in MarsDome. (left) The effect of wheel slippage. (right) Fully autonomous. To verify the GPR was working we buried a 2 cm diameter steel pipe midway into the gravel layer, perpendicular to the rover’s intended direction of travel. We also buried a long piece of aluminum with a vshaped cross section (each side of length 15 cm) in the gravel. Figure 3 (left) shows a straight-line survey with 10 cm horizontal resolution in which these objects have been identified. One can also identify the concrete layer as well as a deeper interface which becomes more shallow as we traverse from left to right. Just after the 20 m mark the rover became stuck in a concave hollow and its wheel slipped significantly before regaining traction. This survey was conducted in a teleoperation mode. Figure 3 (right) shows a 35 m survey with 2 cm horizontal resolution which was conducted completely autonomously. Again, we can identify the concrete layer as well as a deeper interface which is shallower on the left. We did not have an independent groundtruth localization system to determine how well the rover travelled a straight line but it was visually estimated that the rover strayed no more than 1.5 m either side of the intended straight survey line. Thus the lateral (side-to-side) motion was approximately 10% of the survey length, 35 m. As another experiment we took the system to a frozen pond1 and a frozen canal2 to see if we could survey the ice-thickness. The day we conducted the pond experiments there were a few centimeters of crisp packed snow on top of the ice layer. Figure 4 (left) shows a GPR survey of the pond ice thickness as the rover was driven approximately 50 m away from shore. The ice is of fairly uniform thickness near 45 cm (except right by the shore (left)). The horizon1 Grenadier 2 ‘The

Pond in High Park, Toronto Canal’, Ottawa

tal resolution of the GPR data was 2 cm. Figure 4 (right) shows the canal ice-thickness as the rover was driven perpendicular to its length. It was found to be roughly 75 cm thick. The ice experiments were conducted in a teleoperation mode. Figure 5 shows a shorter survey conducted in an office hallway. The concrete foundation and reinforcement rods can be clearly identified. This survey was conducted autonomously with a blue recycling box as the visual target located down the hallway. The GPR horizontal resolution was 2 cm.

6

Discussion

The hardware configuration used here was found to be good in most situations. The parallel arm mechanism used to keep the GPR flat with respect to the ground performed well but did present problems when the terrain was highly concave along the direction of travel. In some cases, the GPR and front wheels remained in contact with the ground but the back wheels were lifted up. This had two negative effects in that the rover was prone to getting stuck and the distance travelled was over-estimated because of spinning wheels. Our suggestion to avoid this is to either have a small trailer for the GPR or to have a swivel joint to allow the system to conform to highly concave terrain. Another situation which revealed a problem with dragging the GPR was on snow. Early in the morning, the GPR slid along the snow surface very nicely but towards noon when the sun begin to melt the snow slightly, it would easily build up in front of the GPR and eventually cause the rover to get stuck on a mound of packing snow. This could likely be avoided by extending the GPR’s skid plate into more of a skitype shape, possibly waxing the bottom of the unit

Subsurface Survey on Frozen Pond

Subsurface Survey on Frozen Canal

4

4

x 10

x 10

3

3

Top Surface of Ice

Top Surface of Ice

2

5

4

2

2

Bottom Surface of Ice (~45 cm thick)

6

10

Bottom Surface of Ice (~75 cm thick) 1

8 10

Secondary Reflection

0

12

Time [ns]

Time [ns]

1

15 0

20

Secondary Reflection −1

−1

14

25

16

30

18 20

−2

−2

−3

3

6 9 12 15 18 Distance Along Ground [m]

21

Signal

35

−3

5

10 15 20 25 Distance Along Ground [m]

Signal

Figure 4: Survey of ice thickness on (left) frozen pond and (right) frozen canal. with the appropriate cross-country ski wax. The rover was found to have a tolerable effect on the GPR unit in terms of induced noise. The wireless ethernet communications provided the most noise, particularly when the signal was weak. This is because it is standard practice to retransmit data packets many times if they are corrupted. In the worst case the noise envelope was less than 2% of the analog-to-digital converter’s range. This could be reduced if communications were shutdown during GPR measurements. With fully autonomous surveys there is no reason to maintain communications except for monitoring purposes. The other possibility is to increase the frequency gap between the GPR and communications systems. We also recommend keeping the GPR close to the rover as it is much easier to turn and the noise levels are not significantly elevated by doing so. Although odometry is notoriously bad at providing long-term localization information, particularly on loose terrain, we found it was adequate for creating GPR surveys if used mainly as a measure of distance travelled. Most of the time only slight slippage occurred which was evenly distributed along the GPR survey. Thus, a stretch factor could be applied to correct for this simple kind of slippage. Occasionally, a more serious form of slippage would occur, for example, the rover getting stuck. Figure 3 (left) shows the effect this has on the GPR survey. However, unless all four wheels are slipping by the same amount, it is possible to detect this kind of slippage after the fact. The rover used here has open differentials joining the left-right pairs of wheels. Thus, when it is stuck and slipping, only one wheel from each pair is turning while the other remains relatively still, providing a detection criterion. We later hope

to incorporate a traction control technique to detect such problems and overcome them. An inertial sensor or texture-based visual odometer (similar to an optical mouse) would complement the wheel-encoder odometry at estimating distance travelled. We found that using a visual landmark was a good way of getting orientation information. The simple colour segmentation method for feature detection described here did work. It showed that a combination of visual servoing and odometry was reasonable to conduct straight-line GPR surveys with fine spatial resolution. However, more sophisticated visual servoing techniques would certainly improve this approach. For example, in one trial the blue target was lost entirely which caused the rover to deviate beyond recovery from its intended path. We hope to replace the method described here by a model-based pattern identification method of recognizing both the location and orientation of a simulated lander in future experiments. Extending the technique described here to full threedimensional GPR surveys requires additional work. It should be possible to do this once our vision system is improved. With only one visual landmark (e.g., a lander), the most practical “grid” shape may not be the typical rectangle. For example, spiraling outward from the landmark might provide a simpler problem to solve in terms of navigation and control while still yielding good scientific data. Using more visual landmarks will certainly improve the localization and may help with occlusion problems which we have yet to deal with in this work. We conveniently chose terrains free of rocks, trees, crevasses and other hazards. A truly autonomous GPR system should certainly be capable of detecting these and adapting its sample grid correspondingly.

Subsurface Survey in Office Hallway

4

x 10

Top Surface of Floor

2

1

4 0.5

Time [ns]

6 8

Rebar in Concrete Foundation

0

10 −0.5

12 14

−1

16 18 20

−1.5

1

2 3 4 5 6 Distance Along Ground [m]

7

Signal

Figure 5: Survey in office hallway showing reinforcement bars in concrete foundation.

7

Conclusion

We have presented our experiences to date with a rover and on-board ground-penetrating radar. We found that simple straight-line GPR surveys could be conducted autonomously using a method based on odometry and visual servoing. Preliminary experimental results were presented on realistic outdoor terrains. We find that it is realistic to automate the process of GPR surveying. We also find that particularly for planetary exploration, GPR offers an opportunity to do useful science with existing rover technology. Acknowledgments We would like to thank David Redman of Sensors & Software Inc. for advice on how to proceed with this work. The research was funded in part by the Center for Research in Earth and Space Technology and the Natural Sciences and Engineering Research Council of Canada.

[4] J L Davis and A P Annan. Ground penetrating radar for high-resolution mapping of soil and rock stratigraphy. Geophysical Prospecting, 37:531–551, 1989. [5] Alex Foessel, Dimi Apostolopoulos, and William Whittaker. Radar sensor for an autonomous antarctic explorer. In Proc. SPIE , Mobile Robots XIII and Intelligent Transportation Systems, volume 3525, pages 117–124), January 1999. [6] Alex Foessel, Sachin Chheda, and Dimitrios Apostolopoulos. Short-range millimeter-wave radar perception in a polar environment. In Proceedings of the Field and Service Robotics Conference, August 1999. [7] J A Grant, B A Campbell, and A E Schutz. A rover deployed ground penetrating radar on mars. In Proceedings of the IEEE Conference on the Geophysical Detection of Subsurface Water on Mars, 2001. [8] J A Grant, A E Schutz, and B A Campbell. Exploring the martian highlands using a rover-deployed ground penetrating radar. In Proceedings of the IEEE Workshop on the Martian Highlands and Mojave Desert Analogs, 2001. [9] Andrew Mishkin, Jack Morrison, Tam Nguyen, Henry Stone, Brian Cooper, and Brian Wilcox. Experiences with operations and autonomy of the mars pathfinder microrover. In Proceedings of the IEEE Aerospace Conference, volume 2, pages 337–351, March 21-28 1998. [10] Gary R Olhoeft. Ground penetrating radar on mars. In Proceedings of the 7th International Conference on Ground Penetrating Radar, pages 387–392, May 2730 1998. [11] Liam Pedersen. Science target assessment for mars rover instrument deployment. In Proceedings of the 2002 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 817–822, Lausanne, Switzerland, September 2002.

References

[12] Liam Pedersen, Michael Wagner, Dimitrios Apostolopoulos, and William Whittaker. Autonomous robotic meteorite identification in antarctica. In Proceedings of the 2001 IEEE International Conference on Robotics and Automation, Seoul, Korea, May 2126 2001.

[1] Steven A Arcone, Norbert E Yankielun, and Edward F Chacho. Reflection profiling of arctic lake ice using microwave fm-cw radar. IEEE Transactions on Geoscience and Remote Sensing, 35(2):436–443, 1997.

[13] C Samson and K Ait-Abderrahim. Feedback control of a nonholonmic wheeled cart in cartesian space. In Proceedings of the 1991 IEEE International Conference on Robotics and Automation, Sacramento, USA, April 1991.

[2] Deepak Bapna, Eric Rollins, John Murphy, Mark Maimone, William Whittaker, and David Wettergreen. The atacama desert trek: Outcomes. In Proceedings of the 1998 IEEE International Conference on Robotics and Automation, Leuven, Belgium, May 1998.

[14] G Visentin, M Van Winnendael, and P Putz. Advanced mechatronics in esa’s space robotics developments. In Proceedings of the 2001 IEEE/ASME International Conference on Advanced Intelligent Mechatronics (AIM), pages 1261–1266, Como, Italy, July 8-12 2001.

[3] Robert Beck and James Osborn. Three dimensional migration and forward modelling of ground penetrating radar data. Technical Report CMU-RI-TR-91-12, Robotics Institute, Carnegie Mellon University, Pittsburgh, PA, June 1991.

[15] Wayne Zimmerman, Robert Bonitz, and Jason Feldman. Cryobot: An ice penetrating robotic vehicle for mars and europa. In Proceedings of the IEEE Aerospace Conference, volume 1, pages 311– 323, 2001.