Self-localization of autonomous
vehicles in horticulture environments
Yadvir Singh
August 12, 2015
Supervisor(s): Roy Bakker (UvA)
Inf
orma
tica
—
Universiteit
v
an
Ams
terd
am
of the solutions is to use autonomous vehicles. These vehicles can lift some repetitive tasks o↵ human hands. Tasks include for example crop spraying or automatic harvesting. To perform these tasks, the robots have to be able to navigate through greenhouses and fields autonomously and therefore needs to be aware of its position. In this thesis, di↵erent tracking solutions are compared, of which a RFID grid based approach is concluded to be best promising on both cost and durability. Experiments showed that RFID performance was sufficient in outdoor environ-ments and test runs concluded that the proposed tracking technique is able to navigate a robot through a preset course.
1 Introduction 5
1.1 Outline . . . 5
2 Related work 7 3 Platform & Environment 9 3.1 The environment . . . 9
3.2 The platform . . . 10
4 Tracking techniques 11 4.1 Odometry . . . 11
4.2 Active beacons . . . 11
4.2.1 Sound based tracking . . . 12
4.3 Landmark based tracking . . . 12
4.3.1 Single landmark based tracking . . . 12
4.4 Map based tracking . . . 13
4.4.1 WLAN mapping . . . 13 4.4.2 RFID . . . 13 4.5 Comparison . . . 14 4.6 Conclusion . . . 15 5 Implementation 17 5.1 Algorithm . . . 17 5.1.1 Start . . . 18 5.1.2 Tag detection . . . 18
5.1.3 Control and Navigation algorithm . . . 18
5.2 Points of interest . . . 19 5.2.1 Sample run . . . 19 5.3 Software . . . 20 5.3.1 Memory . . . 20 5.4 Hardware setup #1 . . . 21 5.4.1 Wiegand . . . 21 5.5 Hardware setup #2 . . . 22 6 Experiments 23 6.1 Indoor experiments . . . 23 6.1.1 POI design . . . 23 6.1.2 Test run . . . 24 6.1.3 Speed test . . . 25 6.2 Outdoor experiments . . . 26 6.2.1 RFID performance . . . 26 6.2.2 Test run . . . 28 6.3 Active beacons . . . 29
7 Conclusion 31 7.1 Future work . . . 32 7.2 Reader recommendations . . . 32
Introduction
Autonomous vehicles in horticulture are mainly intended to take over repetitive work done by humans. This work may include automatic crop spraying or functioning as a caddy to support workers in transporting goods. To enable a robot to perform these tasks, it has to be aware of its position. Di↵erent techniques pose di↵erent solutions for tracking robots, one of which is GPS. This system provides a fairly cheap solution to track vehicles, however it struggles in indoor environments, such as greenhouses, where metal structures prevent good reception. This research project will concentrate on finding a tracking solution that can perform well in horticulture environments.
The Hortimotion robot (see figure 1.1) developed by Jacco Boersen is used as a use-case. Di↵erent tracking solutions for the Hortimotion robot were tried, but none proved to be ideal. Firstly, GPS repeaters were considered. These repeaters are ground based beacons that provide better accuracy and reception than satellite based GPS. Although providing better reception, its setup and licensing costs were too high to be economically viable. Another technique was tested which used a cable to guide the robot through the greenhouse. Again, cost was an issue and it became clear that the laying of cables involved too much labor.
Demands to the solution were: the solution must be economically achievable in the sense that no expensive hardware must be used, maintenance should be kept to a minimum and lastly, it should be able to function in horticulture environments.
The research question that is answered in this thesis is: Which tracking technique proves to be a solution for the Hortimotion robot?
1.1 Outline
Chapter 2 will discuss some automated vehicles and their tracking techniques that are designed to work in horticulture environments. Next, chapter 3 introduces the Hortimotion prototype robot on which the tracking technique is implemented and tested. Also, a typical environment in which horticulture robots operate, is discussed. Chapter 4 discusses some of the most common techniques to track robots. Techniques are compared on the basis of ease of deployment, accuracy, durability and maintenance requirements. Finally, one technique is considered to be a solution for tracking robots in horticulture environments, and will be further investigated. Chapter 5 presents the implementation and workings of the proposed tracking technique. Chapter 6 discusses the experiments conducted to test RFID performance in outdoor environments and overall tracking performance of the proposed tracking technique. Chapter 7 concludes the research and gives recommendations for future work and for better hardware.
Related work
Tracking of autonomous machinery in horticulture and agriculture environments is not new [7]. This chapter discusses some implementations of di↵erent tracking techniques used in horticulture vehicles.
Figure 2.1: Hot water pipes used as tracks [18].
The use of mechanical guidance is described by Gan-Mor et al [10] and Michelinit et al [1]. In this technique the robot uses tracks to travel in between pathways. The use of tracks eliminates the need for expensive and com-plex navigation solutions. Although simple, this technique is still used today. The robot developed by Rafiq et al [18] uses a similar technique. This robot uses hot water pipes (see figure 2.1) that function as guiding rails. The pipes that are used are intended to transport hot water and al-ways occur as a pair of pipes that run parallel to each other.
Figure 2.2: The AURORA robot [8].
Figure 2.3: The FITROBOT [11].
The general purpose AURORA [14] robot developed at the University of M´alaga used sonar sensors to navi-gate in greenhouses. The sonar setup consist of: four Short-Range Digital (SRD) sensors, two Mid-Range Dig-ital (MRD) sensors and four Mid-Range Analog (MRA) sensors with adjustable orientation. Besides these dis-tance sensors, a video camera is mounted at the front of the vehicle to provide a live video stream for the user.
The FITOROBOT [11] is designed for automatic spraying ac-tivities. It is an example of combining several di↵erent sensor techniques. It activities include: spraying, pruning, and crop transport The robot uses seven ultrasonic sensors to detect dis-tances from obstacles, one magnetic compass for determining ori-entation, two incremental encoder (odometry) to track angular speed and position, one radar to measure linear speed and finally security sensors to prevent collisions with crop rows, walls and other obstacles.
Platform & Environment
This chapter will discuss a typical environment in which autonomous horticulture vehicles operate in. Next, the the platform on which the Hortimotion is build, is presented.
3.1 The environment
The Hortimotion robot is designed to be operated in horticulture environments. A field trip to Van Reeuwijk Fruit and Flowers located in Abbenes was conducted to see in person which kind of conditions the robot is to be operated in. The trip showed that greenhouses di↵er in size but have 20 pathways on average. Along the edge of each pathway, racks with fruits are stationed above the ground level. Pathways are not always open ended on both sides. There might be several pathways that are closed on one end. A small corridor runs perpendicular to pathways in order to connect them to each other. See figure 5.2 for a sample greenhouse layout.
In the sample layout, the robot is only able to make turns at either ends of pathways (path 1, 2, 4) when it enters the corridor, or it can rotate about its own axis at closed ends (path 3). Using its sensors, the robot is able to move between pathways without colliding with fruit racks. It can continue driving forward until either it needs to turn left, or right, or turn around. Only at these points the location needs to be known, while this is not mandatory while driving through pathways. Figure 3.1 shows one of the greenhouses visited during the trip. This particular greenhouse is build on top of a grass field. The surface consists of short grass and dry soil. Not all greenhouses are build on this surface. Concrete and stone tiles are also used as surfaces.
To feed water to fruits, an automatic irrigation system is installed. Electric power (for water pumps at 24V) is available only at selected points, often placed at 3 to 4 pathway intervals. Beside these electrical points, no overhead wires are available. Tracking equipment that requires power can thus only be installed at selected points. A passive system that requires no power is therefore desirable.
Figure 3.1: Visited greenhouse in which strawberries are grown. The Hortimotion robot will drive along these pathways completing di↵erent tasks.
3.2 The platform
As the Hortimotion robot is still in development, its construction is mainly geared towards prototyping. Di↵erent prototypes have been developed over the past years and the current version is based on an Arduino Mega. Two wheels are mounted on either side and can be controlled independently by two motor controllers. This enables the robot to turn about on its own axis. To prevent tipping over, two additional wheels are mounted underneath the robot.
Power is provided by two 12V lead batteries connected in series to provide 24V. The entire robot is controlled by the Arduino Mega. It is responsible for controlling the motors (using PWM to control the motor controllers) as well for interpretation of sensor data. Sensors include two ultrasonic sensors that are used for collision avoidance. Corrections made using these sensors make the robot able to move through pathways using fruit racks as reference points. Secondly a 433 MHz module is available for wireless communication. Lastly, an emergency button is in-stalled on top of the robot to make the robot brake and stop immediately.
Although the Arduino is satisfactory for prototyping, it is not the ideal platform for the final release product. The limited storage capacity can become an issue when storing data of large maps. Besides memory shortage, processing power is limited. This can cause problems when working with navigation algorithms that require significant processing power to compute navi-gation plans in acceptable times. Alternatively, a Raspberry Pi can used to do computational intensive task and use its GPIO pins to communicate with the Arduino or directly to the motor controllers and sensors.
Tracking techniques
Di↵erent tracking techniques o↵er di↵erent solutions. Some of the most common techniques are discussed in this chapter after which they are compared on cost, deployment, accuracy, durability and maintenance requirements.
4.1 Odometry
Odometry is widely used as a tracking instrument for mobile robots. It is inexpensive, has good accuracy over a short distance and allows for high sample rates [5]. Odometry works by sampling wheel revolutions. A single increment can be mapped to x meters linear displacement by the corresponding wheel. Although this method is fairly trivial, it is susceptible to errors. Errors might for example be created by wheel slippage, in which case wheel rotation will not accurately be translated to the linear displacement. These errors can propagate over time which will de-crease accuracy. Errors created by odometry can be categorized in two groups [6]: systematic errors and non-systematic errors. Systematic errors are caused by imperfection of the robot, for example unequal wheel diameters. Non-systematic errors are caused by interaction with the floor. These errors include wheel slippage and bumps. Due to these errors, odometry is often used in conjunction with other sensors.
A technique similar to odometry is used by systems based on inertial navigation [5]. These systems use accelerometers and gyroscopes to measure rate of rotation and acceleration. This data can be mathematically integrated to find the position. As with odometry, this system is susceptible to error propagation (finite accuracy when integrating) over time. Both systems are not suited for long-term position tracking but can be used in short distance operations where recalibration points are available.
4.2 Active beacons
Active beacon systems consist of active transmitters and a receiver. These systems can provide high accuracy in conjunction with high sampling rates at the cost of a high purchase price [5]. Ac-tive beacon systems can be subdivided in to two categories: trilateration and triangulation based. Trilateration is based on distance measurement to known beacon sources. This technique usually consist of at least three or more transmitters and one receiver [5]. The transmitters are located at fixed positions while the receiver is mounted on a mobile robot or vice versa. Using time-of-flight data, the receiver is able to calculate the distance to the transmitter. Global positioning system (GPS) is an example of trilateration.
Triangulation requires at least three or more active beacons. The receiver is mounted on the mobile robot and should be able to rotate freely around the longitudinal axis. Transmitters are mounted at fixed positions. The receiver then actively searches for a signal from one of the transmitters. When found, the angle of rotation is recored. Using this data, the position and the heading of the robot can be determined [5].
4.2.1 Sound based tracking
Sound based tracking is based on an active beacon system where beacons are placed at fixed locations. These beacons are configured to send sound pulses at pre-stated intervals. The receiver on the robot is capable of receiving these pulses and calculates its position using trilateration. According to [4] accuracies of +/- 1 cm can be achieved in an area of 4⇥4⇥9 meters. Besides its good accuracy, no specialist hardware is required which cuts the cost. The main disadvantage of sound based tracking is the amount of beacons needed to cover a relative small area. Accuracies described in [4] are achieved using eight beacons to cover an area of 4⇥9 meters. As an average greenhouse is 10 times larger, the amount of nodes needed and the corresponding setup cost are economically no longer viable. Deployment of these beacons is fairly easy but requires accurate measurements between nodes [4] to increase tracking accuracy.
4.3 Landmark based tracking
Landmark based tracking is based on recognition of geometric shapes. These shapes often have a fixed position and should be chosen carefully to maximize recognizability (for example: contrast to the background) [5]. The characteristics of these shapes need to be known by the robot prior to operation. When the robot recognizes a object, it can determine its position relative to the recognized object. Once landmarks are placed, they require little to no maintenance. The prob-lem of finding a landmark can be simplified by assuming that the current position is known (for example using odometry) [5]. This reduces the area in which the robot has to search for a given landmark. Landmarks themselves are further categorized into artificial landmarks and natural landmarks. Natural landmarks can be described as objects that are already part of the envi-ronment. Artificial landmarks are specially designed to maximize recognizability. Varying light conditions can cause trouble recognizing landmarks as shapes cannot be distinguished anymore. Lighting conditions might vary by changes in the environment, weather conditions or simply by loss of sunlight in the evening.
4.3.1 Single landmark based tracking
Experiments of single landmark based tracking [3] have shown that positional errors of less than 10 cm are achievable. The position is calculated by measuring the distance to the landmark at two di↵erent points using a stereo vision camera. In order to recognize the landmark and do the range measurements, a clear view of sight is needed. Any obstruction, for example by humans passing by or displacement of environment objects (crop growth for example) can prevent readings and therefore the position estimation. This makes the system less durable in horticulture environments.
To capture stereo vision images, di↵erent types of cameras can be used. Depending on the setup, cost can vary from less than §100 to upwards of §1500 for native stereoscopic cameras [12].
of the robot. While in operation the robot actively generates a map of its local environment and this subset is compared to the map previously stored [15, 5]. If a match is found, the robot is able to determine its location. A good reference map should consist of objects that are easily detectable and stationary to maximize the chance of an unique match. The main advantage of map-based navigation is that it can work without modifying the environment. The demands to the reference-map imply that there are enough stationary objects [5] and minimal changes to the operating environment. These demands form the greatest disadvantages of the technique.
4.4.1 WLAN mapping
WLAN (wireless local area network) mapping is a technique that is categorized as a map-based navigation technique. This mapping technique generates a map that consist of RSS (Radio signal strength) measurements of WIFI access points at di↵erent positions. Afterwards, map-matching in conjunction with the current measured signal strength is used to approximate the location (interpolation used in between measurement points). The main advantage of this approach is the limited use of dedicated hardware. WIFI access points that are already installed can be used [15] which reduces cost. The lack of accuracy and durability form the greatest disadvantage of this system. According to [15] accuracies of 1.5 m can be achieved when using enough access points. These values are not sufficient for the given problem, as turning space at the end of aisles might be smaller than 1.5 m in which case the robot initiates a turn at the wrong position. Besides its accuracy, the horticulture environment can cause issues as WLAN mapping su↵ers just like any map-based navigation from changes in the environment (in this case RSS values change). Any major disruption of the radio map will require regeneration of the radio map, which makes maintenance mandatory.
4.4.2 RFID
As RFID (Radio frequency identification) becomes more and more popular, its entry into the robot tracking field has followed consequently. Work of [19] showed RFID being used in a map based navigation system. The technique uses passive RFID tags placed in the environment which are read using a RFID reader with multiple antennas pointing in di↵erent directions. A list of detected tags with the number of detections represents a snapshot of the current location, which can be compared to previously stored points.
Another approach was taken by [20]. The proposed system uses RFID tags placed at ground level. These tags are mapped to a physical location. The robot is equipped with a RFID reader that detects tags as it drives over them. Tags are placed in a grid like structure in which each tag covers a small square area that symbolizes a unique position. These tags can be hidden underneath carpets or wallpapers to seamlessly incorporate into rooms.
4.5 Comparison
The previous sections discussed some methods of tracking autonomous vehicles. The following section will discuss the strengths and weaknesses of each of these methods.
WLAN mapping has low accuracy [15] and is vulnerable to changing RSS values caused by environment changes. In horticulture environments changes may include: equipment movement, natural changes due to crop growth and human or animal presence in the environment. This problem can be partially overcome be regenerating the radio map. This procedure requires sub-stantial more maintenance than comparable techniques. For this reason WLAN mapping will not be a candidate for further investigation.
Single landmark based tracking o↵ers good accuracy [3] in combination with low maintenance requirements. In order for successful position estimation, a clear view of sight is needed. This condition can not be guaranteed in horticulture environments, where dirt or overgrowth of plants can obstruct the view. This makes landmark based navigation unsuitable.
Sound based tracking provides great accuracy at the cost of high setup expenses. Accuracies described in [4] are achieved using eight beacons to cover an area of 4⇥9 meters. As an aver-age greenhouse is 10 times larger, the amount of nodes needed and setup cost are no longer viable. Odometry is one of basic instruments to determine the position of the robot. Providing great accuracy, this technique is also susceptible to error propagation [5] which can accumulate over time causing great inaccuracies of position estimates. These errors can partially be described to wheel slippage (non-systematic errors). The horticulture environment often consist of dirt pads that can become slippery or are uneven and therefore pose the greatest risk of non-systematic errors. As such, odometry is not reliable enough to function one its own, but can be used with other tracking techniques.
RFID based tracking can be separated into two techniques. Firstly, RFID is used as a map based tracking system [19]. Being a map based technique, it su↵ers from environmental changes. The second technique proves to be a better solution. Ground tags provide at the spot location information [20], and can be placed at specific positions where location awareness for the robot is desired. This is important because the Hortimotion robot does not have to be aware of its position all the time. Hardware cost are acceptable when opting for a general purpose reader in combination with passive RFID tags.
count for performance in indoor environments but rather performance in outdoor environments, especially greenhouses. Sound based tracking [4] WLAN map-ping [15] Single landmark tracking [3] Odometry [5] RFID [19, 20] Cost 3 3 2 4 3 Deployment 2 1 3 4 2 Accuracy 4 0 2 1 2 Durability 1 0 1 3 2 Maintenance 3 1 3 0 4 Total score(20) 13 5 11 12 13
Table 4.1: Tracking techniques awarded points on performance in di↵erent categories.
4.6 Conclusion
RFID based navigation poses the best solution. Firstly, because the costs of a basic RFID setup are bearable. RFID readers and tags come in di↵erent variants operating in roughly three major frequency categories. These frequencies include 125 KHz, 13.65 MHz and 865-868 MHz. High frequencies are able to cover greater distances than their counterparts and are therefore used in applications that require minimum read ranges of several meters. High frequency readers tend to be more expensive than medium and low frequency readers.
One of the most common used frequency today is 13.65 MHz. This frequency is often found in credit card sized tags used for access control. Among others, thanks to mass production a basic setup can be bought for less than §20. Similar to 13.65 MHz, the 125 KHz frequency is used for access control and identification purposes, however now a days it surpassed by the medium frequency band.
Secondly, passive RFID tags are virtually maintenance free. They contain no moving parts and are battery free which enhances life expectancy. Because passive tags don’t require exter-nal power sources, their deployment in electricity-free rural environments is possible without deployments of an electric grid.
Thirdly, when opted for the grid structured approach, tracking points can be freely chosen in the sense that the parts of the grid where position does not have to be known are not tracked. The grid structured approach allows RFID tags to be placed at points where the robot is due to perform an action other than driving forward. This can for example include: turns, stops and robot specific actions. Because greenhouses often consist of several pathways side by side, RFID tags can for example be placed at the end of aisles to indicate turns, or in pathways to turn on sprinklers, rather than tracking the robot constantly.
Because of low cost, good life expectancy and the adaptive tracking abilities, the grid based RFID approach proves to be best for the Hortimotion robot in the horticulture scene.
Implementation
This chapter discusses the implementation of a RFID grid based tracking system. Firstly the tracking algorithm is explained followed by an overview of the written software and tag placement strategy. Finally two di↵erent hardware approaches are presented.
5.1 Algorithm
The following flowchart illustrates the workings of the implemented system. Each color repre-sents di↵erent stages that are discussed individually.
Start position Move forward Perform action Sensed RFID tag? Card in database? Control algorithm Return error Return path data no yes no yes
5.1.1 Start
A map, containing the ID’s of tags and their physical location should be loaded into the robot prior to operation. At start, the robot can be placed anywhere inside the greenhouse. The robot is not aware of its position at that moment. At start, the robot starts with its default behavior: moving forward. Using its ultrasonic sensors (and possibly other collision avoidance sensors), the robot continues moving forward and simultaneously scans for RFID tags using the reader mounted underneath.
5.1.2 Tag detection
While moving forward, the RFID reader is actively scanning for tags. If no tag is found, the robot returns to its default behavior. This loop is unbroken till a tag is detected upon which the tag’s ID is compared against stored entries in the internal database. If a match is found, corresponding data about the current path is retrieved and returned. If no match is found, an error code is returned. This approach di↵ers from the RFID map based tracking [19] as the robot is not aware of its position all the time, which is not mandatory as there are limited positions where the exact location is needed to be known. A more closer match is the grid based approach by [20]. As limited points are of interest, tags are not placed in a grid like fashion but are only placed at points where location matters, such as turning or task points.
5.1.3 Control and Navigation algorithm
The data returned by the database query is processed by the control and navigation algorithm. This algorithm decides what action the robot should perform next, based on the data received. Actions can include: stop, turn left, turn right, turn 180 degrees, switch on or o↵ sprinklers, or continue moving forward.
Separation of the tracking and navigation algorithm, makes this technique more flexible. The programmer is free to use or implement any navigation algorithm of his liking.
RFID tags are used to indicate actions other than default and are placed at ground level or below. Only at these ”Points of interest” (POI) location awareness is necessary. Each POI has to be identified uniquely as it represents one unique location. To di↵erentiate tags, their fixed ID is used. Figure 5.2 shows an example of POI placement. In this figure, pathways 1,2 and 4 are all open ended. POIs are hereby placed at both ends where the robot cannot continue forward and thus has to turn. Pathway 3 consists of one closed end. A POI is placed at the closed end to indicate a 180 degree turn. Other than pathway ends, POIs can also be placed inside pathways (for example POI #9 in figure 5.2). These POIs indicate actions other than turns, for example: turn on and o↵ sprinklers or change of pesticide.
5.2.1 Sample run
The control algorithm is programmed to guide the robot through each pathway in the sample layout of figure 5.2. Firstly the robot is positioned in pathway 1 and starts moving in the direction of the arrow. As the RFID reader does not sense any tags yet, the robot continues moving forward (its default behavior). As the robot moves over POI #2, the reader picks up ID information of the tag and the control algorithm will tell the robot to make a 90 degree right turn. Next, the control algorithm tells the robot to move forward and eventually it will cross POI #4. Again the control algorithm will tell the robot to make a 90 degree right turn. As the robot moves through path 2 it detects POI #9. This POI does not indicate directional changes but activation of sprinklers mounted on the robot. Next the robot encounters POI #3. At this point the robot is instructed to turn around. On its way back, POI #9 will be detected again. The control algorithm is responsible to either ignore (in which case it has to remember scanned POIs) the tag or respond with an appropriate action. At POI #4 the robot will turn 90 degrees right once again. At POI #6 the robot will make another right turn and enters pathway 3. As this pathway is closed on one end, POI #5 will indicate a 180 degree turn heading back towards POI #6. Lastly the robot turns right at POI #8 and finishes its traversal at POI #7.
5.3 Software
The software that handles communication between the reader and, storage and retrieval of POIs, is developed to run on the Arduino platform. The Hortimotion robot uses an Arduino Mega which is an extended version of the most commonly used Arduino: the Uno. The Mega provides more input and output pins and holds more memory compared to the Uno. Although di↵erent, programs might be interchangeable between di↵erent Arduino boards. To make the software more easy to use and add ability to make future adjustments, the code is written as a library that can be included in a sketch (synonym for an Arduino program).
For communication with the RFID reader, the Wiegand-Protocol-Library-for-Arduino is used [13]. This library handles communication using the Wiegand protocol. To simplify use, the written library contains a single function that will check if the reader has read any tag, if not, the control algorithm will be notified. If a tag is read, its ID is compared to entries in the database. Information about path number and direction (to indicate which end of the pathway) is returned upon successfully finding the ID. If not found, a notification is returned to inform the control algorithm that a tag is detected but not found in the database.
The next subsection will discuss memory management of POIs.
5.3.1 Memory
Points of interest are stored as struct objects. Firstly, the path number is stored to di↵erentiate pathways. Pathways can have two entries, one at each end. To determine at which end the POI resides, a variable called pathDirection is used. As POIs can consist of multiple cards (see experiments section) a variable called cardCount is used to keep track of the amount of tags. Lastly an array to store identification numbers of cards associated with the POI is used.
Struct objects can be stored in two di↵erent ways. They can either be defined directly in the sketch by creating struct objects on start and storing them in an array. When using this approach, it is easy to modify POI structs because they are part of the code and can visually be altered. The main disadvantage is that POI structs will have to be hard coded into the sketch and are initialized each time the program runs.
The second approach uses the non-volatile memory of the Arduino to store struct objects. This memory is known as the EEPROM (Electrically erasable programmable read-only memory) and can be written to and read from while the Arduino is in operation. This type of memory has a live expectancy of 10.000 write operations per cell. The Arduino Uno o↵ers 1024 bytes of EEPROM storage while more expensive models such as the Arduino Mega o↵er upwards of 4096 bytes. The standard EEPROM library (available in Arduino IDE 1.6.1 or newer) is used to handle read and write operations to the EEPROM. Both write and read operations need an index at which the data is to be written to or read from. This requires the programmer to keep track of the current index. The main advantage of this method is that POIs that have been stored, do not have to be initialized at runtime and will still be available after power-cuts and even when di↵erent sketches are loaded. The main disadvantage is that POIs stored in the EEPROM, can not be altered as easy as hard coded POIs.
The written library takes care of all read and write operations to the EEPROM. On initializa-tion, the last 4 bytes of the memory are reserved for administration purposes. Two bytes are used to keep track of current memory index. This value represents the first address that is free and can be written to. Two additional bytes are reserved to keep track of how many POIs are stored . The limited storage capacity of the Arduino can become an issue when storing all POIs of large greenhouses. In the sample greenhouse represented in figure 5.2, 9 POIs were used. At
5.4 Hardware setup #1
The first hardware approach consists of a Promag GP-25 RFID reader mounted underneath the robot (see figure 5.3) in conjunction with passive RFID tags placed at ground level or below. The opted Promag GP-25 reader is a general purpose 125 KHz reader intended for identification purposes. It has an advertised read range of up to 25 cm and is enclosed in an watertight case which is beneficial in horticulture environments as moisture is common due to watering of plants and rainfall. Readers vary in price from less than §20 to upwards of §1000 for industrial grade readers. At §80 the Promag reader falls in the lower end of scale and is chosen for research purposes only. Although better readers are available, test are conducted to determine reader performance and its capability to function in the final release version of the robot.
To accommodate di↵erent security systems, di↵erent types of interfaces are provided by the chosen reader. Firstly, the industry standard Wiegand protocol is supported. This protocol is often found on commercial RFID/Magnetic readers. Besides Wiegand, support for serial communication is also available through the RS232 serial interface. Wiegand is the preferred choice over the serial connection as there is no voltage di↵erence between the reader and the master device(Arduino) that receives information. This saves the need for including a voltage level converter. Distance from the reader to the ground surface is fixed at 5 cm to allow for better recognition of ground tags and to avoid collision with debris from the ground.
As the reader works at the specific frequency of 125 KHz, tags are used which operate at the same frequency. To keep cost down, standard (EM4001 ISO) ID access cards are used. These cards come factory loaded with a unique ID that is also been printed on each card to make identification more easy. See figure 4.2 for a visual representation.
Figure 5.3: Side view of the robot with the RFID reader mounted underneath.
5.4.1 Wiegand
The reader provides three connection wires for the Wiegand protocol, namely DATA 0, DATA 1 and a common ground wire. When a ’0’ is sent, DATA 0 is pulled to low voltage while DATA 1 stays at high voltage. Consequently when a ’1’ is sent, DATA 0 stays high and DATA 1 is pulled down. When no data is sent both wires stay high. The standard voltage for the Wiegand protocol is 5V which makes it the preferred protocol to use as it operates at the native Arduino voltage, independent of the voltage that is supplied to the reader (5-12V).
5.5 Hardware setup #2
RFID tags come in two variants, namely passive and active tags. Passive tags are solely powered by energy from radio waves send by the reader. Therefore passive tags do not need any internal power source which makes them durable and maintenance free. Because power is drawn from radio waves, range is often limited and will depend on the signal power emitted by the reader and the material surrounding the tag.
Active tags fall in the opposite category. Unlike passive tags, active tags are battery powered and can transmit data at pre-set intervals or transmit upon receiving a signal from the reader. The independent power source makes for greater range by generating more powerful response waves, but inevitably batteries will have to be replaced which requires maintenance over time. Although battery powered tags provide better range, they can cost up to 1000 times more than passively powered tags.
Tags have to be placed at each end of pathways, therefore cost quickly rise beyond viability. As an alternative, an Arduino in combination with reed sensors can be used to ”imitate” active tags. Reed switches are triggered when they are near to a magnetic field. Normally open (NO-open) reed switches are used to ensure that electricity is only flowing when a magnetic field is present, opposed to normally closed switches. The reed switch is connected in between the power supply and the Arduino. When no magnetic field is present, the Arduino is cut from power as the reed switch will stay open. This also prevents unnecessary power consumption when no robot is near the tag. The magnetic field required for the reed switch is generated by a Ferrite magnet that is mounted underneath the robot, similar to the RFID reader. Whenever the robot passes over the reed switch, current will start flowing and the Arduino will power up and transmit the small message which will be received by the robot. To let the robot be aware of its position, a small message containing the path number and direction are transmitted wirelessly. Trans-mission is handled by 433 MHz modules using the virualwire library [16]. A receiver module is mounted underneath the robot to pickup signals transmitted from any “active” tag. Noise does e↵ect these modules so a 16 bit CRC checksum is used by the library to verify integrity of the message. See figure 5.4 for the schematic representation.
Experiments
This chapter discusses some of the experiments done on the two implementations discussed in the previous chapter.
6.1 Indoor experiments
6.1.1 POI design
Prior to tests using the Promag reader, a generic (ACR122U based) USB RFID reader was used for exploring RFID capabilities. A small scale platform was built that could be moved freely (to simulate the robot) to which the reader was mounted underneath. The first test consisted of tags laid in a straight vertical line at 10 cm intervals. While moving the platform over the line, it became clear that the reader has a very limited field of view. This translates to cards only being read if they are 1 cm o↵ the center axis of the reader. Similar performance is noted with the Promag reader at 2 cm variance.
To cope with the limited field of view, tags can be laid in di↵erent formations. The best formation proved to be diagonally, see figure 6.1. The diagonal shape o↵ers the most coverage with the same amount of cards compared to cross or linear laid tags. This is important because POIs should be able to be read from any direction. For example, in figure 5.2 POI #6 can be read from 3 di↵erent angles. Either the robot has turned around at POI #5, or it is coming from either POI #4 or POI #8.
Although the diagonal shape o↵ers good coverage, a better reader with an wider field of view will reduce the need of multiple tags.
6.1.2 Test run
To test the functionality of the proposed tracking system in small scale, a L-shaped course was built indoors. POIs consisted of 6 cards that each covered a square area of 0.25m2. Two POIs were located at both ends of the course, indicating a 180 degree turn. A third POI was used to indicate the 90 degree turn in the L-shape. As the robot has no capability of sens-ing direction, turns were made ussens-ing timed wheel rotation. Lastly, reference points were used to measure o↵set of the robot against the center of POIs. Measurements were taken from the reference points to the wheels of the robot. At start, the robot was placed at the starting point located inside the course. When instructed, the robot started its control-algorithm and moved forward heading west. As the robot detected the first POI, it initialized the 180 de-gree turn. The last scanned POI was remembered to avoid confusion as the robot may detect the same POI again during its turn. As the robot headed east, the second POI instructed the robot to make a 90 degree right turn after which the robot stops at the last POI. Marker points at each POI were used to measure the o↵set of the robot against the center of POIs. To make measurements more convenient, the robot was programmed to stop at each detected POI. The following map shows three separate runs indicated by lines colored green, blue and red.
Pathway, 1 meter wide POI center Reference point Runs 1, 2 and 3 ”L” shaped course outlay
210 cm 175 cm x axis y axis Starting point
Figure 6.2: Map illustrating the L-shaped course of the indoor test. The lines colored green, blue and red depict the traveled path by the robot on each run.
Figure 6.3: Two POIs placed 8 cm apart
the same position on all three runs. When moving to the second POI, the heading of the robot varies slightly on run 1 compared to runs 2 and 3, resulting in an standard deviation = 2.83 (measured in direction of movement). Standard deviation increases to x = 8.83 at the last POI. This shows that without any knowledge of direction, the robot will slowly drift o↵ course. This error will propagate over time and therefore, POIs are not reliable enough on their own. The proposed sys-tem should be used in conjunction with other sensor information. Ultrasonic sensors or electronic compasses can for example correct the robot as at drifts o↵ center. Apart from drift, a problem arises when a POI is not detected. Without any other data, the robot will not be aware of its failure to detect a POI. This can be solved by using odometry sensors. Distances between POI can be measured exactly and compared to odometry data. Using the combined data, the robot can “expect” when to detect a POI. POIs can also function as recalibration points to compensate for odometry errors.
6.1.3 Speed test
In order to test the read speeds of the chosen reader, speed test were conducted. The test setup consisted of
two POIs (consisting of 6 cards) placed at a distance of 8 cm from each other, to see if the reader is able to detect both POI at di↵erent speeds(see figure 6.3). The first test consisted of the robot moving at 25% (0.21 m/s) of its capable speed.
In all 5 test runs, the reader was able to detect both POIs. Next, the speed was increased to 50% or 0.48 m/s. Again the reader was able to read both POIs at all 5 runs. Although tested at 50% of its speed capacity, the robot will operate at lower speeds in commercial use. Also, POIs will generally be placed more than 8 cm apart. The test concludes that that the reader is certainly able to read tags at commercial operating, and even higher, speeds.
6.2 Outdoor experiments
6.2.1 RFID performance
As RFID tags will be placed in outdoor environments they will meet di↵erent surfaces in di↵erent conditions. These variables might a↵ect RFID performance and hence a test setup is build to test RFID performance in these varying conditions.
The test setup consisted of the Promag GP25 reader in conjunction with standard EM4001 ISO 125 KHz tags. The reader was held horizontally (facing downwards) by two stands on either side. The height of the reader can be adjusted in steps of 15 mm. To conveniently test di↵erent surface materials a small plastic container was used (see figure 6.4). The container was filled up to the inner marking of 1100 ml with the material to tested. Cards were buried at 4 and 6 cm, both flat and at a 45 degree angle. Each test was conducted at two di↵erent voltage levels. Firstly tested at 4.95V, the standard output voltage of the Arduino, to see performance when no additional power source is available. Secondly a 9V power source was used to see if increasing the voltage has any e↵ect on the performance. Both sand and garden soil were tested as these are common surface material found in greenhouses and finally a limestone brick was used to test the e↵ects of more dense materials.
Rain was also simulated by pouring 460 ml of water in the box. On average, there are 20 days at which more than 20 mm rainfall is measured in the Netherlands [17]. This equates to 20 liters per m2. The surface of the box measures 15.2 x 15.2 cm = 0.0231 m2. This equates to: 0.0231 x 20 = 0,46 L = 460 ml rain collected by the container.
Each test was conducted in the following order:
Figure 6.4: Plastic container used to test di↵erent surface materials. 1. Fill container with 1100 ml of material.
2. Insert tag at 4/6 cm.
3. If tag is sensed, raise reader by 1.5 cm and repeat step.
4. Repeat step 3 at 9V.
4.95V 4.95V water 8.80V 8.80V water 5 6 7 8 5 5 8 8 depth in cm 4.95V 4.95V water 8.80V 8.80V water 0 2 4 0 0 5 5 4.95V 4.95V water 8.80V 8.80V water 5 6 7 8 5 5 8 8 4.95V 4.95V water 8.80V 8.80V water 0 2 4 0 0 5 5 depth in cm Garden soil 6 cm 4.95V 4.95V water 8.80V 8.80V water 0 2 4 6 0 0 6.5 6.5
Garden soil 4 cm 45 angle
4.95V 4.95V water 8.80V 8.80V water 0 2 4 6 0 0 6.5 6.5 Sand 4 cm 45 angle 4.95V 8.80V 0 2 4 6 0 5.5 depth in cm Limestone brick
Figure 6.5: Bar plots showing the maximum depth at which a RFID tag is still detectable under di↵erent materials/condition.
The 4 cm sand test shows that the tag is still detectable at both 4.95V and 8.80V. The voltage increase clearly shows increased reader performance as the tag is detected at 8 + 4 = 12 cm from the reader. At both voltage levels, water seems to have no e↵ect on the reading performance as readings are equal to their dry counterpart. This may be explained by the displacement of water. As water is poured on the material, not all of it is concentrated on top or near the buried tag, but is rather equally spread throughout the container. This results in little rain water present between the tag and the surface.
Buried at 6 cm the low voltage of the Arduino is not capable of reading the tag anymore. At 8.80V the reader is able to read the tag but su↵ers from the increased depth of the card. The total distance to the reader is 6 + 5 = 11 cm compared to 12 cm from th 4 cm sand test at 8.80V. Thus, although distance from the reader to the surface has reduced significantly, the total distance from reader to tag has only dropped by 1 cm. Both the garden soil test at 4 cm and 6 cm show similar results. This is due to similar properties of both materials (both are less dense materials).
Next, tags were tested at an angle of 45 degree. At 4.95V the reader is unable to read the tag. At 8.80V the tag is detectable with the reader at 6.5 cm from the surface. At an angle of 45 degree, the reader has more trouble reading the card compared to the flat laid card used in the 4 cm sand test. This result concludes that for optimal recognition, tags should be placed in level with the surface.
Finally, a test was conducted with a limestone brick to see how solid materials e↵ect read performance. The limestone had a thickness of 4.9 cm and the reader was placed at 5.5 cm above the brick. At 4.95V the reader was unable to detect the card. At 8.80V the tag was detected but the performance is notably less compared to the less dense materials.
6.2.2 Test run
The second test run was conducted outside. Two POIs were placed at 210 cm apart. The first test consisted of both POIs placed at ground level (see figure 6.6). By the second test, both POIs were buried about 4 cm be-low ground level. Each test consisted of 10 runs. The following table shows the results. At each run, detec-tion of each POI is marked bypas recognized or⇥ to indicate failure. Run POI Ground level 4.95V POI Buried 12V 1 p p p⇥ 2 ⇥p p p 3 p p p p 4 ⇥ ⇥ p p 5 p ⇥ p p 6 p p p p 7 p p p p 8 p p p p 9 p p p p 10 p ⇥ p p Table 6.1: Results outdoor run
second run concludes that the proposed system is capable of operating outside.
6.3 Active beacons
The two main components of the second hardware implementation consist of the reed switch and the Arduino. Both components were tested separately. First the reed switch was tested using a circular shaped ferrite magnet to find its detection range. The test showed that 8 cm was feasible, although a more powerful magnet can increase this value. The power should however be chosen carefully, as too powerful magnets will disrupt electronics and motors.
Secondly the Arduino was tested. Tests showed that when the reed switch is activated and current starts flowing, a 1 second gap is present between the Arduino receiving power and execution of the code. Even if the robot is traveling at 1 km/h (0,28 m/s) the Arduino will not be able to finish its booting process and no message will be sent. This problem can partially be overcome by placing a capacitor in the circuit. The capacitor will supply the required power when the reed switch is open (when no magnetic field is present). Now that the message can be sent to the robot, the robot itself would have covered a distance of 28 cm (1 km/h) the aruidno will still be booting. This delay in reception results in a wrong position estimation, but can be rectified by calculating the distance traveled based on the robots current speed.
Thirdly power consumption was looked upon. This is important as power consumption must be down to a minimum to preserve battery power and reduce maintenance. As the Arduino micro controller (Atmel Atmega328) is not the most efficient solution, the MSP430 by Texas Instruments is used as a reference. The MSP430 has an average power consumption of 0.35mAh [9]. Furthermore, this controller works at 3.3V and can be operated from a coin cell battery. The following values are used in the life expectancy calculation:
• 0.35 mAh consumption by the MSP430 micro controller • 20 mAh consumption by the 433 MHz transmitter
• 150 mAh at 3.3V is supplied by a coin cell battery (CR2032) • an avrage of 10 transmissions per day
The life expectancy is than calculated by: 0.35+203600 = 0.0057 mA per second is consumed by the controller and transmitter. Multiplying this value by 10 gives the consumption per day: 10⇥ 0.0057 = 0.057 mA per day. This equals to 150
0.057 = 2.6⇥ 102 hours or 110 days of use.
This means that batteries have to be replaced roughly every 4 months which raise maintenance costs. Due to these maintenance requirements this approach proves to be not the ideal solution for tracking the Hortimotion robot.
Conclusion
This research found and evaluated a satisfactory tracking system for the Hortimotion robot. Firstly, available techniques were compared. Sound based tracking looked fairly promising with its high accuracy, but the amount of nodes needed to cover a complete greenhouse made it eco-nomically not viable. WLAN mapping provides the benefit of using widely available networking equipment. The biggest drawbacks of this technique is low accuracy and its vulnerability to environmental changes that causes RSS values to change followed by a mandatory remapping. Single landmark based tracking provided good accuracy, low maintenance and depending on the setup, o↵ered low costs. Not able to guarantee clear sight of view and requirements of sufficient illumination make this technique not reliable enough. Lastly two di↵erent RFID based tracking implementations were examined. As with any map based approach, environmental changes cause disruption of the stored map and therefore the map matching technique proved not to be the ideal solution. The second technique studied used a grid layout in which every cell is occupied by a RFID tag. An adapted version of this technique proved to be best on cost, durability and maintenance. As the Hortimotion robot does not have to be aware of its position all the time, tags can be placed only at turning/task points.
Two di↵erent hardware implementations were designed. The first implementation consisted of a RFID reader mounted underneath the robot. Tags are placed at points of interest (POIs) where the robot is supposed to turn or perform a task. Whenever the robot pases over tags, the ID of the card is read and the navigation algorithm dictates the next action based on the perceived ID. The second implantation is functionally equal to the first one. Rather than passive tags, an active RFID tag is simulated by an Arduino in combination with a reed switch that is activated by a magnet mounted underneath the robot. When the reed switch closes due to the presence of the magnet, the Arduino is powered on, and a small message is sent to the robot containing information about the path at which the tag is placed.
Next tests were conducted to see how RFID performs in outdoor environments and to see how the proposed tracking technique holds up. At first it became clear that the chosen reader has a limited field of view. To solve this issue, POI were constructed from several cards laid in a diagonal fashion.
Stationary benchmarks proved that tags can be read at an distance of 8 cm when buried 4 cm in both garden soil and sand. Simulated rain seemed to have no e↵ect on the readers performance, but reading performance su↵ered badly from cards positioned at a 45 degree angle. This test concluded that RFID performance is sufficient in outdoor environments for the proposed tracking technique.
Finally the tracking technique was tested both indoor and outdoors. The indoor run showed that POIs are able to guide the robot through a course. As the robot moved from POI to POI, it started to drift o↵ course. The drift concluded that the proposed tracking technique cannot be used without a secondary sensor, preferably odometry. When using odometry, the POI can function as both turning and task points and, recalibration points for odometry data. The
out-door test concluded that the proposed tracking system is able to function outside, even if tags are buried at 4 cm.
Overall the proposed system proved to be a good solution for tracking the Hortimotion robot. RFID performance was sufficient in outdoor environments and POIs placed at turning points were able to navigate the robot through a predefined course. The proposed system is however not able to function on its own. A combination with other sensor data is necessary to compensate for drift and failure to detect POIs.
7.1 Future work
In the proposed solution only ID’s of RFID tags are stored. As the robot has the tendency to drift over time, the unique ID of RFID tags can be used to recalibrate the driving direction. As POIs consist of multiple tags their distance to the center tag can be used to correct drift. Along with storing ID’s, their distance to center tag can also be stored. Whenever the robot scans a tag, the POI information along with the o↵set is retrieved and the o↵set can be used to correct drift.
Tags can also be laid in a cross like fashion. Although less surface is covered compared to diagonally laid tags, it can help when the robot needs to turn. When turning, the robot has to remember which line of the cross it has read when scanning the POI. The robot can than start rotating till it reads a tag that is associated with line perpendicular to the line firstly scanned, resulting in a 90 degree turn. This solution is however unnecessary when using sensors, like magnetic compasses, to determine orientation.
Other than laying tags at ground level or below, tags can also be mounted to fruit racks or other stationary objects. The distance from the robot to these tags will generally be greater compared to ground based tags. Therefore readers with a minimum read range of 1 meter are recommended. See section 7.2 for recommendations.
7.2 Reader recommendations
Although the Promag reader held up good in di↵erent tests, it is far from the ideal reader. The following readers have better specifications for the final release product:
1. TowiTek TWT2015, §169, 125 kHz, https://www.conrad.nl/nl/towitek-rfid-lange-afstandslezer-twt2015-module-12-vdc-191527. html 2. UHFSKY UHFSKY-0702,$135 - 220, 860MHz - 960MHz, http://www.alibaba.com/product-detail/passive-long-range-5m-rfid-reader_60096948022. html?spm=a2700.7724838.35.1.KcrEmF 3. Generic, 125 kHz, §88,10, http://ebayitem.com/161313221170
4. N-POLE, NP-U83 UHF RFID reader,$150 - 180, 902-928MHZ,
http://www.alibaba.com/product-detail/Passive-long-range-3-6M-UHF_60278292376. html?spm=a2700.7724838.35.1.PGpdQh
5. Keway K100D,$130 - 150, 125 KHZ,
[1] GM Acaccia, RC Michelini, RM Molfino, and RP Razzoli. Mobile robots in greenhouse cultivation: inspection and treatment of plants. In Proceesings of the the 1st International Workshop on Advances in Service Robotics, Bardolino, Italy, volume 1315, 2003.
[2] Adafruit. Microsd card breakout board+. https://www.adafruit.com/products/254. [3] Abdul Bais, Robert Sablatnig, and Jason Gu. Single landmark based self-localization of
mobile robots. In Computer and Robot Vision, 2006. The 3rd Canadian Conference on, pages 67–67. IEEE, 2006.
[4] Jan D Bjerknes, Wenguo Liu, Alan FT Winfield, Chris Melhuish, and Coldharbour Lane. Low cost ultrasonic positioning system for mobile robots. Proceeding of Towards Au-tonomous Robotic Systems, pages 107–114, 2007.
[5] Johann Borenstein, Hobart R Everett, Liqiang Feng, and David Wehe. Mobile robot positioning-sensors and techniques. Technical report, DTIC Document, 1997.
[6] Johann Borenstein and Liqiang Feng. Measurement and correction of systematic odometry errors in mobile robots. Robotics and Automation, IEEE Transactions on, 12(6):869–880, 1996.
[7] Stephen Cameron and Penelope Probert. Advanced Guided Vehicles: Aspects of the Oxford AGV project. World Scientific Publishing Co., Inc., 1994.
[8] J. Manuel Gomez de Gabriel. Teleoperation of the aurora mobile robot for greenhouse spraying. http://gomezdegabriel.com/projects/old-projects/ teleoperation-of-the-aurora-mobile-robot-for-greenhouse-spraying/.
[9] ElPerfecto.com. How much power does msp430 / ti launchpad use? http://www. elperfecto.com/2011/01/08/how-much-power-does-msp430-ti-launchpad-use/. [10] Samuel Gan-Mor, Benjamin Ronen, S Josef, and Y Bilanki. Guidance of automatic vehicle
for greenhouse transportation. In International Conference on Greenhouse Technologies 443, pages 99–104, 1996.
[11] R Gonz´alez, F Rodr´ıguez, J S´anchez-Hermosilla, JG Donaire, et al. Navigation techniques for mobile robots in greenhouses. Applied Engineering in Agriculture, 25(2):153, 2009. [12] Point Grey. Bumblebee2. http://www.ptgrey.com/
bumblebee2-firewire-stereo-vision-camera-systems.
[13] JP Liew. Wiegand-protocol-library-for-arduino. https://github.com/monkeyboard/ Wiegand-Protocol-Library-for-Arduino.
[14] Anthony Mandow, Jesus Manuel Gomez-de Gabriel, Jorge L Martinez, Victor F Munoz, An´ıbal Ollero, and Alfonso Garc´ıa-Cerezo. The autonomous mobile robot aurora for green-house operation. IEEE Robotics & Automation Magazine, 3(4):18–28, 1996.
[15] Eladio Martin, Oriol Vinyals, Gerald Friedland, and Ruzena Bajcsy. Precise indoor local-ization using smart phones. In Proceedings of the international conference on Multimedia, pages 787–790. ACM, 2010.
[16] Mike McCauley. Virtualwire library. https://www.pjrc.com/teensy/td_libs_ VirtualWire.html.
[17] Waterservicepunt Nijmegen. Regenwater. http://www.waterbewust.nl/regenwater. html.
[18] Alireza Refigh, Davood Kalantari, and Hamid Mashhadimeyghani. Construction and de-velopment of an automatic sprayer for greenhouse. Agricultural Engineering International: CIGR Journal, 16.
[19] Sebastian Schneegans, Philipp Vorst, and Andreas Zell. Using rfid snapshots for mobile robot self-localization. In EMCR, 2007.
[20] Ricardo Tesoriero, Jos´e Gallud, Mar´ea D Lozano, V´ıctor MR Penichet, et al. Tracking autonomous entities using rfid technology. Consumer Electronics, IEEE Transactions on, 55(2):650–655, 2009.
