POSITIONING AN AUTONOMOUS OFF-ROAD VEHICLE BY USING FUSED DGPS AND INERTIAL NAVIGATION T. Schönberg, M. Ojala, J. Suomela, A. Torpo, A. Halme Helsinki University of Technology, Automation Technology Laboratory Otakaari 5 A 02150 Espoo, FINLAND tel. +358-0-4513300, fax +358-0-4513308 Abstract: The paper deals with positioning system for an autonomous vehicle ARSKA. Localization of the vehicle is based on fusion of internal dead reckoning navigation and periodic absolute position measurements. Fusion is done by using Kalman-filtering technique. Similar kind of approach is used in correcting the heading measurement. This is important because the position error is mostly result of the accumulated heading error. The resulting accuracy depends on the accuracy of the dead reckoning estimation and on the accuracy and frequency of the absolute position measurement. As an absolute position measurement system several alternatives can be used. Two different systems have been used with the vehicle; an external optical measurement device, tachymeter, and different combinations of DGPS. Also a real guarding application will be discussed in the paper. Keywords: Autonomous vehicle, Off-road, Kalman filter, Dead reckoning, DGPS, RDS. 1. INTRODUCTION Mobile robots are becoming a future trend in many areas where manually operated work machines are used today [1 Koskinen 1993]. Some potential applications are material transportation in building sites, forestry, agriculture, open quarries, docks, warehouses and special purpose machines such as harvesters, rock drilling machines, building machines, surveillance applications and cleaning machines for large areas. In this paper a positioning system for an autonomous vehicle ARSKA (Autonomous Robot for Surveillance Key Applications) and a real guarding application, implemented using the vehicle, is discussed. The localization of the vehicle is based on fusion of internal dead reckoning navigation and periodic absolute position measurements. Fusion is done by using Kalman-filtering technique. Similar kind of approach is used in correcting the heading measurement. As position measurement system an external optical measurement device, tachymeter, and different combinations of DGPS (Differential Global Positioning System) are used. The test vehicle used for the results presented in this paper was developed in ESPRIT II project 2483 Panorama as a testbed for developing parts of perception and navigation system. The complete test system includes also a PC-based ground station from which the vehicle is monitored and teleoperated. The test vehicle is a general purpose highly instrumented mobile robot capable to move in natural outdoor environments. It is made from the commercial Honda TRX 350 all terrain vehicle. Throttle, steering, brake and gear switch are controlled by computer. The computer hardware of ARSKA is based on PC cards. The sensing system includes ultrasonic and navigation sensors. First ARSKA s hardware and software are discussed, then the different combinations of GPS systems and DGPS base stations are tested. DGPS systems used
are Trimble and Astech. The DGPS base stations used are a commercial system, broadcasted by radio channels and an Astech base station. The accuracy of the two GPS systems will be compared in the case that the vehicle is moving and not moving. Finally results using ARSKA in a guarding application are presented. 2. THE TEST BED ARSKA The test-bed ARSKA is built on a commercial vehicle Honda TRX 350. TRX 350 is mechanically durable and it is made for rough terrain environments [2 Schönberg 1990]. but is needed for testing of additional systems. As an energy buffers two serially connected batteries of 12 V are used. The 24 V power from the batteries is converted to different voltages with several DC/DC converters. ARSKA can be controlled using a ground station. The basic functions of the ground station are task preparation and monitoring and on-line tele-control of the vehicle [4 Ojala 1991]. In addition to these, ground station uses a servo-controlled optical measurement device, tachymeter, to measure the real trajectory driven by the vehicle. Position measured by the tachymeter can also be sent to vehicle computer in real-time for accurate position estimation. This feature is very useful while testing different positioning systems or vehicle control algorithms. 2.2 Software Fig 1. ARSKA The test-bed is a four wheel drive vehicle, it has five gears forward, one backwards and a centripetal clutch. 2.1 Instrumentation The instrumentation consists of four main parts; the computer, the actuators, the sensors and the radio communication. A PC with 80486 processor is used for the control of the vehicle. The PC is equipped with two Analog & Digital I/O boards and a RS232 additional board. The RS232 ports are used for gyro, modem communication, ultrasound sensors and DGPS. The control of the throttle and steering is done with servo controller cards. The gear shift, brake and turning head are controlled using relays. The feedback from the actuators is measured using potentiometers and inductive displacement transducers. ARSKA is also equipped with ultrasonic sensors that can be used for contour following and obstacle avoidance. [3 Penttinen 1993]. Additionally to the vehicles own 12 V an other generator was installed on it. The capacity of the generator is 24 V and 1 kw, which is more than enough for the PC system, The software of ARSKA can be separated roughly to two levels. The upper level calculates position, accomplishes course following, avoids obstacles, etc. The lower level controls actuators with high frequency. Interface between these two levels is realized by transmitting velocity and curvature setpoints from up to down and low level measurements from down to up. Software is written in C-language on QNX real-time multitasking operating system. 2.3. Localization Localization of the vehicle is based on fusion of dead reckoning and periodic absolute position measurements. Dead reckoning position estimation is calculated at rate of 10 Hz based on vehicle's speed and heading measurements [5 Puputti 1992]. The speed measurement is done measuring the speed of ARSKAs rear wheel, for the heading measurements a fiber gyroscope (Hitachi HOFG-X) and the angle of the front wheels is used. The dead reckoning position estimation gives quite good results on short term basis. Test runs have resulted 1% error on the average. However, absolute position measurements with a proper frequency are required to compensate the accumulating error of the dead reckoning estimation. The dead reckoning estimation and the absolute position measurement are fused by using Kalmanfiltering techniques to provide a corrected estimate. The weights of the measurements used are constants. The DGPS measurements are weighted with zero when there is no measurement and with a 0.015 when there is measurement available. Similar kind
of approach is used in correcting the heading measurement. This is important because the position error is mostly result of the accumulated heading error. The resulting accuracy depends on the accuracy of the dead reckoning estimation and on the accuracy and frequency of the absolute position measurement. 3. GLOBAL POSITIONING SYSTEMS Two different kinds of positioning systems have been used with the vehicle; an external optical measurement device, tachymeter, and differential GPS (DGPS). Tachymeter gives very accurate position measurement at a rate of 1 Hz. By fusing tachymeter measurements with dead reckoning estimations, a continuous position estimation error is <0,3 m. Using tachymeter is however quite cumbersome and needs lot of open space. Thus using this positioning system is meaningful only while testing other subsystems. Even though the accuracy of the DGPS system is much worse, it can be easily utilized anywhere. This is why for absolute positioning mainly DGPS has been used. As DGPS measurements are received at regular intervals, accuracy of the localization depends naturally very much on the vehicle's speed. While testing the abilities and possibilities of the DGPS correction, only one optical gyroscope for heading measurement with no altitude angle compensation has been used. This results as a normally quite accurate dead reckoning system (figure 2), even though heading error can grow dramatically at any time the vehicle is turning (figure 3) [6 Schönberg 1994]. Fig. 3 Heading error in the path following, the maximum error is bigger than 20 meters and the length of the route is 558 meters 3.1 Comparison of GPS measurements There are available very good GPS receivers with an accuracy that surely is enough for positioning of an autonomous vehicle. One example is the best model of Trimble (Trimble Kinematic) that is a realtime instrument with an accuracy of 2 cm. The problem with this receiver is the price of about 60 000$ in Finland. It was of interest to find out which of the less expensive solutions would be possible to use for an autonomous vehicle. Two different GPS receivers were tested, one older model of Astech and a new cheap model of Trimble (SVeeSix PLUS XT DGPS) for which very good accuracy is promised. The Astech receiver does not have any filtering of the data but uses 12 channels and the Trimble receiver has 6 channels and uses filtering to get better measurements. In whole the Trimble GPS system is a low cost easy to use mass product. Astech GPS is more expensive and harder to use, but it gives the user access to the raw data. The comparison of the two GPS receivers was done with two different tests, moving and not moving. In both tests we used differential correction data. [7. Torpo 1994]. 3.2 Measurements when the vehicle is not moving Fig. 2 Accurate path following, the maximum error is about than 2.5 meters and the length of the route is 567 meters As shown in figure 4 and 5 the filtering in the Trimble receiver gives about a decade better results than the results got from the Astech receiver. The measurements are done during a six hour period with a measurement frequency of 1 Hz. According to the measurements it seems that filtering is an easy way to improve the measurements a lot, but as shown in chapter 3.3 this is not the case. The filtering used is made for measuring a stable point.
An other kind of filter is needed while measuring a moving point. Using the kind of filtering that Trimble has while the vehicle is not moving can improve the position information a lot. Trimble. The conclusions are that accurate GPS receivers using filtering for stable points do not do the positioning more accurate but only change the kind of the error. This is the reason why the Astech receiver was chosen for the surveillance robot. Becauce the disturbance in the GPS signal is not predictable and the noise is not white there is of course no way to be certain of the accuracy of the measurement. Even if it may be rare there is a slight possibility for longtime position errors to the same direction that will take the vehicle of the route. This is not actually a problem since there are, as mentioned before, GPS systems available with very high accuracy which is not affected by the added noise. 40 20 Fig. 4 The measurements from Trimble during a 6 hour period. Maximum error is about 10 m. 0-20 -40-60 -80-100 -120-40 -20 0 20 40 60 Fig. 6 Trimble gives quite accurately the shape of the route but recovers very slowly from position error. 40 Fig. 5 The measurements from Astech during a 6 hour period. Maximum error is about 100 m. 3.3 Measurement when the vehicle is moving The measurements while the vehicle is moving show that filtering made for stationary state or slow movement is not usable for positioning an autonomous vehicle moving at a speed of 1m/s. The differential data used for the measurements is broadcasted by radio and received by a Trimble RDS receiver. As shown in figure 6 the Trimble GPS receiver gives quite accurately the shape of the route compared with the measurements of Astech (figure 7). Astech has no filtering so the measurement is not smooth but even if it makes some big jumps from the real route it recovers from the errors faster than 20 0-20 -40-60 -80-100 -120-40 -20 0 20 40 60 Fig. 7 Astech recovers from the errors faster than Trimble.
3.4 Differential GPS The next task was to choose the differential station for the GPS. The first possibility is to use a separate differential station only used for this application. This is the most expensive way to construct a DGPS system. The other possibility is to use a commercial DGPS system. In Finland there are two possibilities to use this kind of system. One of the possibilities is a system that is made for navigating on sea and the other system is transmitted by ordinary radio channels, the Radio Data System (RDS). The system used for navigating on sea sends differential correction data each five seconds and the accuracy is 3m. The receiver for the differential data is quite expensive but less expensive than an own DGPS station for the system. The problem with the previous system is that it is meant for use on sea so it is not promised to work in the direction of land. The RDS promise an accuracy of 2.7 m and the accuracy decreases 10 cm for each 100 km from the base station. The differential information is sent once each 1-2 seconds. The RDS seemed to be the best system for our surveillance application. A Trimble RDS receiver was used for the differential data. The problem with this receiver was that when it looses the radio station it searches through all stations trying to find a new frequency. An Aztech RDS receiver was also tested. The Aztech receiver had the feature of remembering the channels where it has found the DGPS data. If it looses the channel it first tries frequencies that it has used before, this saves a lot of time when the signal has disappeared only for a short time. The channel is lost if the signal disappears for a short time due to an obstacle in the terrain. 4. GUARDING TASK One of the easiest tasks to do with an autonomous robot is the measurement task where information from the environment is collected. Here a guarding task for an army storage area is discussed. Guarding is actually measuring and in this case the vehicle is transmitting a video picture to a base station where the guard is placed. The benefit of the guarding robot is that the human guard can guard the area at the same time as the gate. There can also be more than one guarding robot that sends information to the guard. camera towards for example locks, transmitting the picture to the guard who checks if the lock is okay. The guard at the base station has to confirm that he has seen the picture before ARSKA continues. During our tests the vehicle succeeded in following a net of roads. During the run the vehicle randomly chooses its next goal and aproach it the shortest way. It is also possible to manually choose the goal point. 4.1 Path planning For the vehicle's path planning a preprogrammed map is used. The map actually consists of two different maps a logical map and a geographical map. As can be seen in figure 8 the logical map tells the vehicle where the crossings, turning places and one-way roads are. Fig. 8 The logical map of ARSKA Road Crossing Using the logical map the vehicle does its decisions how to turn in a crossing or making a U-turning. The route is chosen using the priority-first algorithm [8 Sedgewick 1989], where the distance between two points is calculated through all possible crossings and the shortest distance defines the route. First the distances from the first crossing to the nearest crossings are calculated. Next the distances from the new crossings to their nearest crossings are calculated, to these distances the distance to the starting point is added. Continuing like this the distance to every crossing will be known, if there is two possible routes to a crossing the shortest route will be chosen. The geographical map consists of points with a certain distance from each other. The points in the geographical map bound the information in the logical map together. (figure 9). The principle of the guarding robot is that it moves randomly which is hard for a human to do. Because ARSKA is noisy it does not try to surprise an intruder but it stops at certain places turning its
cost about 60 000$ and the gyro that cost 10 000$ will in a few years become much cheaper. Fig. 9 The geographical map of ARSKA Road Crossing Xy-point REFERENCES /1/ Koskinen K, An experimental autonomous land vehicle for off-road piloting and navigation research, 1993, p. 6, IEEE Systems, Man and Cybernetics conference. /2/ Schönberg T, Development and programming of a test vehicle for autonomous applications, Master's thesis (in Swedish), 1990, p. 67 The distance between the points in the geographical map is between 10 and 3 m and the vehicle uses splines to define a route between them, se figure 10. /3/ Penttinen A, Graeffe J, Proximity sensors of Panorama project, VTT/Ins (in Finnish), 1993, p. 20 /4/ Ojala M, Operator-station for an autonomous vehicle, Master's thesis (in Finnish), 1991, p.118 /5/ Puputti J, Development of the piloting system of moving autonomous robot, Master's thesis (in Finnish), 1992, p. 97 Route Crossing Turning points Fig. 10 The route ARSKA moves using splines 5 CONCLUSIONS /6/ Schönberg T, Ojala M, Koskimäki E, Suomela J, Halme A, A Small scaled autonomous test vehicle for developing autonomous off-road applications. 1994 /7/ Torpo A, The development of a guard robot supervising the army storage areas, Master's thesis (in Finnish), 1994, p. 72 /8/ Sedgewik R, Algorithms, Addison-Wesly Publishing Company, USA, 1989, p 660 It is possible to construct an autonomous surveillance robot with the technique available today. The main problem is that the equipment needed for a robust and accurate robot is too expensive for most commercial applications for the robot, however the prices are decreasing all the time. During the tests ARSKA succeeded in following a net of roads quite well. During the run the vehicle uses the priority-first algorithm for choosing the direction to continue in a crossing. Our next steps are to add a separate emergency stop system to ARSKA and take into account the roll and pitch angle of it while measuring the heading and the distance moved. After this improvement it is possible to guard any area with roads of a minimum width of 3.5m. The prices of the components used make it possible to design an autonomous vehicle for guarding applications that pays itself back in about half a year. The most economical way to make the positioning more accurate is to wait for less expensive components. The DGPS system that now