Autonomous Gathering of Livestock Using a Multi
Contents
1. Albright W P Gordon W C Black J P Dietrich W W Snyder and C E Meadows Behavioral responses of cows to auditory training Journal of Dairy Science 49 104 106 1966 2 Z Butler P Corke R Peterson and D Rus From robots to animals Virtual fences for controlling cattle Int J Rob Res 25 5 6 485 508 2006 3 M Culley Grazing habits of range cattle J For 36 715 717 1938 4 Y Guo G Poulton P Corke G Bishop Hurley T Wark and D Swain Using accelerometer high sample rate gps and magnetometer data to develop a cattle movement and behaviour model Ecological Modelling 220 17 2068 2075 September 10 2009 5 K Havstad L Huenneke and W Schlesinger Structure and Function of Chi huahuan Desert Ecosystem New York USA Oxford University Press 2006 6 K H Kwong T T Wu H G Goh B Stephen M Gilroy C Michie and I Andonovic Wireless sensor networks in agriculture Cattle monitoring for farming industries Progress In Electromagnetics Research Symposium 5 1 31 35 March 23 27 2009 7 C Lee K C Prayaga A D Fisher and J M Henshall Behavioral aspects of electronic bull separation and mate allocation in multiple sire mating paddocks Journal of Animal Science 86 1690 1696 2008 8 Phillips LPC241x User Manual 2 edition July 2006 9 N Rutter Time lapse photographic studies of livestock behaviour outdoors on the college farm aberystwyth J Agric Sci2. 71 257 265 1968 10 M Schwager C Detweiler I Vasilescu D M Anderson and D Rus Data Driven identification of group dynamics for motion prediction and control Jour nal of Field Robotics 25 6 7 305 324 2008 11 B Thorstensen T Syversen T A Bjgrnvold and T Walseth Electronic shep herd a low cost low bandwidth wireless network system In MobiSys 04 Proceedings of the 2nd international conference on Mobile systems applica tions and services pages 245 255 New York NY USA 2004 ACM 12 P Zhang C M Sadler S A Lyon and M Martonosi Hardware design expe riences in zebranet In SenSys 04 Proceedings of the 2nd international confer ence on Embedded networked sensor systems pages 227 238 New York NY USA 2004 ACM
3. autonomous methods to monitor animals Thorsten et al 11 presented a low cost wireless communication network system they call the lt Elec tronic Shepherd in order to track sheep during the grazing season Schwager et al 10 and Guo et al 4 used wireless sensor net fr es lt in sy 4 f 0 be af re LE b gt a ais be gas TERA Aj x iae i A w gt h D a gt gt Figure 1 Cow wearing autonomous gathering sensor node works to monitor cattle and calibrated behavioral models for live stock using the gathered data Kwong et al 6 presented a similar system for cattle monitoring with no additional modeling Zhang et al 12 presented a platform to monitor zebras called ZebraNet In the past few years there have been efforts to extend these sensor networks to control free ranging animals Butler et al 2 proposed the use of wireless sensor networks for virtual fencing Lee et al 7 used wireless sensor networks for bull separation Both use similar stimuli as autonomous gathering but achieve dif ferent goals and require different algorithms To our knowledge this paper presents the first work in autonomous gathering using wireless sensor networks 3 Gathering Algorithm In this section we formalize the problem of autonomously gath ering animals provide the intuition on which we based our initial gathering algorithm and present the gathering algorithm consisting of a repetitive cueing loop u
4. cow s ears We measured the max imum output of each speaker at a frequency range of 1 KHz to 10 KHz to be 90dB at a distance of 10cm The entire system is housed inside an OtterBox 3000 case to protect all the components from dirt and water The electrodes and speakers are connected using Switchcraft EN3 connectors The sen sor box is mounted on the neck of the animal with a specially de signed collar that also holds the speakers and electrodes 5 Experiments We performed gathering experiments to validate our algorithm In the first set of experiments we gathered the animals by manu ally applying stimuli using radio control We then performed two autonomous gathering experiments using Algorithm 1 All exper iments were performed on a group of 5 animals Each of the 5 animals was equipped with a sensor node However prior to our radio controlled gathering experiments two of the animals managed to sever the connection between the battery at the bottom of the collar and the sensor node at the top As a result during the ra dio controlled gathering only 3 our of 5 animals were cued and the GPS plots show only these 3 animals However we observed dur ing the experiments that the other 2 animals remained continuously in the group and gathered successfully The experiments occurred on paddocks 7B and 10B on the Jor nada Experimental Range of the United States Department of Agri culture Paddock 7B was used during the radio controlled gat
5. gathered In e the cows showed no long lasting reactions to cueing and did not gather successfully into the corral at all In the cows initially showed a reaction to cueing and started walking towards the corral However after cueing stopped the cows resumed foraging The possible reasons for failing to gather are that the cues in e were not administered frequently enough and in f the cows should have been cued again once they stopped gathering Unfor tunately the radios we used during the experiments performed very poorly offering a range of communication of only a few hundred meters and dropping messages frequently even for distances below 100m As a result we were not always able to continuously cue the animals when desired and had no feedback as to when a cue was actually applied apart from seeing the animals reaction though all applied cues were recorded in log files on the sensor boxes for later analysis This prevented us from always being able to cue the an imals when desired We believe that the bad radio communication a hihuahuan Grassland landscape Mesquite cues were applied shrub is visible in the front every 60 90 seconds tinuously for 20 minutes Semidesert b Jan 29th 5 minutes of aural c Jan 30th Cues were sent in d Jan 31st 30 second long multiple bursts over the course of cues were used with 5 30 second 40 minutes BSD e Feb Ist Cued for 30 seconds f Feb 2nd Cued almos
6. the animal s head location in the direction from which we wish to cue the animal Using the orientation of the cow s head the location of the virtual sound source and the assumption that the cow s ears are spaced by 50cm the cpu computes the running length distances be tween the source and each of the cow s ears From this it computes the attenuation in percent for each ear and their relative phase dif ference in milliseconds This information is computed at 20 Hz and forwarded to the FPGA on the expansion board using a serial pe ripheral interface The FPGA fetches data from a monoaural wave file stored inside the flash It then attenuates and appropriately de lays the signal for each ear to create a stereo directional sound The delay can be set between 0 to 2048 samples for each channel Since the sound file is recorded at 22 KHz this results in a resolution of Algorithm 1 Repetitive Gathering Using Sound SOUND tsound WAIT wait end while Fig dist cuetime each cue no cues time to gather m min sec min 4b 700 5 30 7 13 700 40 30 15 55 900 20 30 16 50 850 48 30 26 no 700 20 60 21 no 3000 240 10 720 no 1700 70 10 210 70 Table 1 Experimental results See Section 5 3 for details 0 05 ms and a maximum delay of 93 ms This delay is sufficient to simulate the sound originating from any direction The signal is then synthesized using the ADC and amplifier and output on two speakers mounted close to the
7. 0 MHz Aerocomm AC4790 radio link The sensor board is equipped with all necessary sensors temperature compass accelerometers and GPS An expansion board provides electric and aural cueing functionality 4 1 Electric Stimuli Subsystem To provide electric stimuli the board is equipped with two drivers and two transformers to generate a high voltage Each trans former s power output is up to 5 00mW but can be scaled by adjust ing the duty cycle The drivers are controlled by two pulse gen erators running inside an FPGA mounted on the expansion board The length of each pulse the distance between pulses and the total number of pulses can be configured independently for the left and right electric stimuli channel from the sensor board using a serial peripheral interface 4 2 3D Sound Subsystem To provide sound cuing the extension board utilizes a FPGA for 3D sound processing and is further equipped with a 2 MB Data flash to store the sound file a 24bit stereo digital to analog sound converter and a stereo amplifier We simulate directionality of the sound by computing the running length distances between the vir tual location of the sound source and the computed locations of the cows ears When cueing the animal aurally the sensor board uses the onboard accelerometer and magnetometer to continuously com pute the yaw pitch and roll of the cow s head The virtual sound source 1s placed at a simulated distance of 10 meters away from
8. Autonomous Gathering of Livestock Using a Multi functional Sensor Network Platform Marek Doniec Carrick Detweiler luliu Vasilescu Dean M Anderson and Daniela Rus t USDA ARS Jornada Experimental Range Las Cruces NM deanders nmsu edu Computer Science and Artificial Intelligence Laboratory Massachusetts Institute of Technology Cambridge MA doniec carrick iuliuv rus mit edu Abstract In this paper we develop algorithms and hardware for the au tonomous gathering of cattle We present a comparison of three dif ferent autonomous gathering algorithms that employ sound and or electric stimuli to guide the cattle We evaluate these algorithms in simulation by extending previous behavioral simulations for cat tle We implemented one of these algorithms and present data from experiments in which cattle were equipped with sensor nodes that allowed cueing with sound and electric stimuli We discuss the min imum requirements for algorithms and hardware for autonomous gathering 1 Introduction Using sensor networks to control livestock is of great interest to the agricultural community The use of sensor networks for live stock allows farmers to monitor and control the herd even when the farmer is far away This can for example help detect illness related behavior improve land management and develop animal behavior models In this paper we focus on using wireless sensor networks to autonomously gather animals Gathering is a r
9. ats until the cow is within of the goal position The parameters t o nq and twait can be random variables to make habituation of the animals to the cues less likely 3 3 Simulation Figure 3 shows the results of a simulation of Algorithm 1 The simulation was based on the behavioral model presented by Schwa ger et al 10 Schwager s model simulates the interaction forces between animals as well as their interaction with the environment using parameters based on previously collected data For the pur pose of this simulation the animals responses to 3D sound cues was assumed to be probabilistic In the simulation the animals walk in the cued direction 50 of the time and do not react to the cue oth erwise Schwager s model describes two forces acting upon the animal between animals and environment to animal We add a third force representing the cueing effects The simulation weights each of these forces by a factor to determine the absolute force acting upon the animal Relative to the Schwager model we use a weight of 1 0 for the interaction force between animals and a weight of 0 4 for the interaction force between the environment and the animal In determining the cueing force no scientific data defines the relation ship between applied stimuli and animal response We specifically choose cueing to have a higher impact than the environment but a lower impact than interactions between the animals giving it a value of 0 6 Th
10. cues if the animals are not moving given proper choices for its parameters tsound twait etc We plan to conduct autonomous gathering experiments using Algorithm 3 in the summer of 2010 7 Conclusion We present a set of algorithms and experiments to gather cattle We provide two sets of stimuli We performed 7 experiments In summary our experiments show that to successfully gather animals we have to take into account many environmental factors such as weather and existing paths in the landscape Further we showed that to reliably gather animals electric stimuli are neces sary in addition to sound cues Sound cues will keep the animal going for most of the time but electrical stimuli can help if the an imal is not reacting to sound This can happen at the beginning of the gather because the animal is resting such as during our second autonomous gathering experiment This could also happen in the middle of a gather such as seen in Figure 4 f Further our ex periments demonstrated the necessity for robust and fault tolerant hardware because of the harsh environmental conditions heat dirt rain and rough handling of the sensors by the animals 8 Acknowledgments We would like to acknowledge the following groups for their fi nancial support Microsoft Research NSF and Smarts MURI We would like to thank the following people for their assistance Eliz abeth Basha Roy Libeau Steven Proulx and Mac Schwager References 1 J L
11. d by the animals Both the reaction to changing weather and the cows adherence to an ex isting path are examples of how the environment plays a major role in controlling cattle 6 Extensions Our results demonstrated the need for a variety of cueing mech anisms and more adaptive mechanisms Based on this we propose two extension of Algorithm 1 1 adding mild electric stimuli and 2 providing more adaptivity 6 1 Repetitive Cueing with Sound Electric Stimuli Algorithm 2 shows repetitive cueing with sound and electric stimuli It extends Algorithm 1 which utilized only sound cues If the cow is not moving after the algorithm plays a sound the al gorithm will apply electric stimuli to the cow In our second experiment we would not have had to intervene were we using Algorithm 2 suggesting that it is a better candidate for gathering In our first experiment it is difficult to predict the outcome had we used Algorithm 2 It might have provided suffi cient stimulus to overcome the effect of the thunderstorm however Algorithm 2 Repetitive Gathering Using Sound and Electric Stim uli while Pcow i Peoal gt do SOUND tsound if COWspeed 0 then SHOCK tshock WAIT wait2 else WAIT wait end if end while Algorithm 3 Adaptive Gathering Using Sound and Electric Stimuli intensity 0 1 while Peow i Peoal gt do if COWsneed 0 then SOUND tsound if intensity gt 0 3 then SHOCK tshock intensity
12. e results of one simulation run can be seen in Figure 3 The cows began at a random location near the middle of the paddock and the simulation ran until the cows gathered We performed a total of 20 such simulation runs and in all simulations the cows were gathered successfully within 45 minutes This validates the theory behind Algorithm 1 4 Platform For the experiments we need a platform capable of gathering basic information such as GPS position and orientation of the cow The platform should offer enough computational power to run all necessary algorithms A radio is also required to allow continuous real time monitoring and remote triggering of the gathering algo rithm Further we need the ability to cue the animals We would like to use two different methods 1 a sound cueing system and 2 an electric stimuli system As a last requirement the platform should be able to run continuously for at least a week to allow for reasonable length maintenance free deployments Figure 1 shows an animal wearing our in house developed sen sor platform outlined in Figure 2 This platform provides the re quired functionality discussed above It is based on the LPC2148 ARM_7 processor 8 running at 60 MHz The processor has 40 KB of on chip RAM and 512 KB of on chip program flash In addition there is a 32 KB FRAM and a mini SD card slot for data storage and logging Communication between different sensor boards and the user is available via a 90
13. end if intensity min 1 0 intensity 0 1 else intensity max 0 1 intensity 0 1 end if WAIT wait end while this is not certain This is because the behavior of cattle in a thun derstorm is difficult to predict In such a case continued electrical stimuli might gather the animal or only increase its stress levels without positive effect The study of this relationship remains fu ture work 6 2 Adaptive Cueing with Sound Electric Stimuli During the radio controlled gatherings we observed that the animals responded better to longer aural cue windows of approx imately 30 seconds or longer Algorithm and Algorithm 2 should provide cues of this length to ensure that the animals respond How ever prior observations of gathering by humans suggests that the animals need the voice command only when they stop moving or begin moving in the wrong direction The animals did not need to be cued by the humans when their heading towards the gathering goal was approximately correct Part of the reason is that the ani mals make a network of roads on their paddocks They know where these roads are and tend to follow roads once on them Therefore once the animals are on on these paths we are less likely to need to provide stimuli Based on these observations we present Algorithm 3 as an ex tension to Algorithms 1 and 2 Algorithm 3 adapts cueing fre quency and intensity to the animal s behavior The outermost loop is responsible
14. for stopping the cueing once the animal has reached the goal location The first of the two if statements are responsible for checking if the cow is moving It is important to note that COWsneeq 1s the component of the cow s actual speed that points towards the goal position If the cow is not moving COWsneeqd 0 we first cue the cow with sound If the cow has not been moving for a few cycles intensity gt 0 3 then we also apply electric stimuli to the cow The strength of the electric stimuli increases with the value of intensity The variable intensity always remains in the range 0 1 Its value increases when the cow is not moving and decreases when the cow is moving This means that when the cow does not move the strength of the electric stimulus gradually increases Therefore to prevent excessive electrical electric stimuli to the animal we upper bound intensity If on the other hand the cow is moving not COWspeed 0 then we do not cue the animal and decrease the value of intensity with a lower bound of 0 This algorithm ensures that the animal is cued only if it is not already moving towards the goal If it stops moving we always first use sound cues If the animal habituates to the sound and does not respond to the cue the sound cues are reinforced with increas ing electric stimuli Both cues stop as soon as the animal starts responding An advantage of Algorithm 3 is that it will automat ically perform long bursts of
15. hering experiments and can be seen in Figure 4 b It has the shaped of an isosceles triangle with sides of approximately 1500 m west and east and 1300 m south The corral is located at the northern tip Paddock 10B is diamond shaped with side lengths of approximately 2100 m shown in Figure 4 g The corral is located in the southern corner The vegetation is a Chihuahuan Semidesert Grassland 5 It is rel atively bush free but offers the occasional obstacle to the cows in the form of Yucca trees and mesquite Figure 4 a shows a sample of the vegetation 5 1 Gathering with Radio Control We performed a total of 5 gathering experiments with radio con trol in paddock 7B between Jan 29th and Feb 2nd 2009 In prepa ration for the experiments we drove a boom truck to the middle of the paddock about 800 m south of the corral as seen in Figure 4 b The boom truck served as the observatory and base station for the researchers during the experiments The goal was to gather and move the animals to a corral located at the northern end of the pad dock The animals are usually moved to this location when manual gathering is performed For all 5 experiments the animals initial start locations were between 700 m and 900 m from the goal loca tion The cows were equipped with the sensor nodes the day before the experiments started ensuring they were not influenced by hu man presence when the experiments began We performed one gathering on each of
16. iment the animals did not initially respond to sound cues While the gathering algorithm was running autonomously we man ually used the radio link to apply electrical stimuli to the animals Specifically we applied this stimuli in 100 ms bursts every 20 sec for approximately a 5 min period After this period the cattle began moving and we did not interfere further with the autonomous aural gathering algorithm 5 3 Results Table 1 summarizes the experimental results for gathering with radio control Figures 4 b e and autonomous gathering Fig ures 4 g h The second column gives the animals approximate distance from the goal at the beginning of the gathering experiment The third column is the time from the beginning of the first sound cue to the end of the last sound cue during the experiment The fourth column is the length of each individual sound cue applied The fifth column gives the total number of cues applied during the experiment averaged across the 3 animals For the autonomous gathering cues were applied every 20 seconds 0 05 Hz resulting in the total seen in the fifth column The last column gives the total time from the beginning of the experiment first cue applied un til the animals gathered A no means the animals did not gather successfully in that experiment The GPS plots for all gathering with radio control experiments are shown in Figure 4 In experiments b c and d the cows were successfully
17. outine husbandry practice that requires animals to be gathered to a specific location for example a corral Currently when a producer performs gather ing he determines the animals locations and drives out to manually gather the animals If the paddock and herd are large the cost of gathering is significant We aim to automate this process In this paper we present algorithms and hardware to au tonomously gather cattle We present experiments demonstrating the use of our hardware to gather cattle both manually using remote radio control and autonomously based on a predetermined sched ule This paper is organized as follows Section 2 reviews wireless sensor network applications in animal monitoring and control In Section 3 we present the initial algorithm for autonomous gather ing and verify it in simulation Section 4 presents the sensor node hardware platform used for experiments Section 5 describes the experiments and results Section 6 discusses experimental results and presents extensions to the first algorithm Section 7 concludes 2 Related Work An early experiment performed in 1966 by Albright et al 1 demonstrated the use of sound to move animals The experiment strapped large tape recorders to one animal of the herd The tape recorder was monaural preset to trigger at a set time using voice commands that the cows were previously habituated to from routine husbandry practices More recently sensor networks have provided
18. pose that changes in weather introduce changes in the environmental force upon the animal In the case of our failed gather we presume that the force introduced by the light ning storm simply outweighed the force exerted by aural cueing The utilization of electric stimuli could possibly overcome this and is introduced as a possible extension in Section 6 1 In the second autonomous gathering experiment the cows did not initially move when the gathering algorithm started cueing them Only after we cued the cows with electric stimuli using the radio link did the cows start moving However once they were moving no additional electrical stimuli were necessary and the cows gather directly into the corral Looking at the forces acting upon the animal the cows behavior during the second autonomous gathering experiment suggests that the state of the animal presents a fourth force that possibly needs to be overcome In this model the ini tial cueing with only sound did not present a strong enough force to overcome the force exerted by the animal s desire to remain for aging However once we increased the cueing force acting upon the animal by using electric stimuli the animals started gathering This changed their state and thus the force exerted by the animals state making it possible to gather the cows without further electric stimuli It is also worth noting that the walking pattern visible on the GPS plot is actually a path frequently utilize
19. sing only sound 3 1 Problem Statement We assume there are n cows with positions P R for i 1 n We are given a goal position P R where we want the cattle to gather The problem is to move all the animals to the goal location such that P P lt for some E R 3 2 Repetitive Cueing with Sound We designed the gathering algorithm based on the experience and intuition of animal scientists who routinely work with livestock The goal is to develop an automatic herding algorithm that is low stress and natural for the animals Humans gather animals by riding or driving behind the group and giving voice commands The gath ering algorithm presented here tries to simulate this experience Storage FPGA Aerocomm AC4790 l i 900 MHz Radio i Low power LTC1733 Battery Circuit 2MB Data Flash FPGA 3D sound Stereo Amplifier 2 Channel 24 bit DAC 2 Channel MOSFETs Transformers for Electric Shock Extension Board Figure 2 Overview of the platform used during gathering experiments The extension Figure 3 Simulation of Algorithm 1 board right side is usually turned off to save power It is only activated when the sensor See Section 3 3 for a description board left side decides to cue the animal Algorithm 1 outlines a gathering algorithm utilizing only sound cues Every iteration provides an aural cue to the cow with the loop condition ensuring that the algorithm repe
20. t con g Aug 11th First autonomous h Aug 12th gathering pauses in between Second au tonomous gathering Figure 4 b f GPS plots of 3 cows during radio controlled gathering experiments g h GPS plots of 5 cows for au tonomous gathering experiments The boom truck s location is marked by a star and cues are marked by yellow dots b f and green diamonds g h The corral is located at the northern b g and southern g h end of the paddock In all experiments the cows were foraging when the first cues were applied The cows gather successfully in a b c and h and did not gather in d e and g was a significant reason for why we were not able to gather the ani mals in e and f We believe that the use of autonomous gathering algorithms can help overcome this problem since radio communi cation is not essential once the autonomous gathering algorithm is started The GPS plots for the two autonomous gathering experiments can be seen in Figures 4 g and h In the first experiment the cows started gathering successfully However during the gather ing the weather changed significantly We observed stronger winds as well as clouds and lightning Since cattle do not behave pre dictably in storms 3 9 we attribute this failed gather to the weather change In Section 3 3 we explained how we simulated three forces acting upon each animal between animals environment to animal and queuing force We pro
21. the following 5 days at which point we re moved the boxes For each of the 5 experiments we began the experiment between 7 00 h and 10 45 h local time In starting the experiment we trav eled with the base station equipment to the boom truck at least 30 minutes before starting the experiment We chose a 30 minute in terval to prevent possible influence of the scientists on the animals Once the experiment started we used the base station to send aural cuing commands to the animals over the radio 5 2 Autonomous Gathering We performed two autonomous gathering experiments on Au gust 11th and 12th 2009 in paddock 10B Both experiments uti lized Algorithm 1 presented in Section 3 We stopped the algorithm once the animals were gathered successfully or after 4 hours in case gathering was not successful In preparation for the experiments we drove the boom truck serving as the observatory and base station to the middle of the north east fence of the paddock as seen in Fig ure 4 g On August 1th we released the cows at the north end of the pad dock at approximately 13 00 h We set the sensor boxes to initiate gathering at 14 30 h on August 1th and 8 45 h on August 12th On August 11th the cows were approximately 3000 m away from the goal location while on August 12th when the gathering algorithm started the animals were located in the middle of the paddock ap proximately 1700 m away from the goal location During the sec ond exper
Download Pdf Manuals
Related Search