ROBOTIC TANKETTE FOR INTELLIGENT BIOENERGY AGRICULTURE:DESIGN, DEVELOPMENT AND FIELD TESTS1

Marco F. S. Xaud∗, Antonio C. Leite†, Evelyn S. Barbosa2∗, Henrique D. Faria2∗,Gabriel S. M. Loureiro2∗, Pal J. From∗

∗Norwegian University of Life Sciences,Faculty of Science and Technology,

P.O. Box 5003, NO-1432, As, Norway.

†Pontifical Catholic University of Rio de Janeiro,Department of Electrical Engineering,

Postal code 22451-900 Rio de Janeiro RJ, Brazil.

mafernan@nmbu.no, antonio@ele.puc-rio.br, evelynsoaresb@poli.ufrj.br,

hd.faria@poli.ufrj.br, gloureiro@poli.ufrj.br, pal.johan.from@nmbu.no

Abstract— In recent years, the use of robots in agriculture has been increasing mainly due to the high demandof productivity, precision and efficiency, which follow the climate change effects and world population growth.Unlike conventional agriculture, sugarcane farms are usually regions with dense vegetation, gigantic areas, andsubjected to extreme weather conditions, such as intense heat, moisture and rain. TIBA - Tankette for IntelligentBioEnergy Agriculture - is the first result of an R&D project which strives to develop an autonomous mobilerobotic system for carrying out a number of agricultural tasks in sugarcane fields. The proposed concept consistsof a semi-autonomous, low-cost, dust and waterproof tankette-type vehicle, capable of infiltrating dense vegetationin plantation tunnels and carry several sensing systems, in order to perform mapping of hard-to-access areas andcollecting samples. This paper presents an overview of the robot mechanical design, the embedded electronicsand software architecture, and the construction of a first prototype. Preliminary results obtained in field testsvalidate the proposed conceptual design and bring about several challenges and potential applications for robotautonomous navigation, as well as to build a new prototype with additional functionality.

Keywords— Mobile robots, Agricultural robotics, Embedded electronics, Sensors Integration.

Resumo— Nos ultimos anos, o uso de sistemas roboticos na agricultura tem aumentado principalmente devidoa alta demanda em termos de produtividade, precisao e eficiencia, causadas pelos efeitos das mudancas climaticase pelo crescimento da populacao mundial. Ao contrario da agricultura convencional, fazendas de cana de acucarsao geralmente regioes com densa vegetacao, areas gigantescas, e sujeitas a condicoes meteorologicas extremas,como calor, umidade e chuva intensa. TIBA - Tankette for Intelligent BioEnergy Agriculture - e o resultadopreliminar de um projeto de P&D cujo objetivo e desenvolver um sistema robotico movel e autonomo capaz derealizar diversas tarefas agrıcolas em fazendas de cana-de-acucar. O conceito proposto consiste de um pequenoveıculo tipo tanque, de baixo custo, a prova de poeira e agua, capaz de se infiltrar em densas vegetacoes emtuneis de plantacao, carregando diversos sensores e dispositivos para realizar tarefas de mapeamento em areasde difıcil acesso e coleta de amostras. Este trabalho apresenta uma visao geral do projeto mecanico do robo,uma descricao da arquitetura de software e eletronica embarcada e a construcao de um primeiro prototipo.Resultados preliminares obtidos em testes de campo validam o projeto conceitual proposto e levantam variosdesafios e potenciais aplicacoes para navegacao autonoma do robo bem como para construir um novo prototipocom funcionalidades adicionais.

Palavras-chave— Robos moveis, Robotica agrıcola, Eletronica embarcada, Integracao de sensores.

1 Introduction

In recent years, the use of robots in agriculture or“agbots” has been growing significantly due to thehigh demand for increased productivity and preci-sion, which was brought about by climate changesand their effects on the environment (Billingsleyet al., 2008). Moreover, the growth of the world’spopulation has motivated farmers to seek an in-crease in the efficiency of food production, for in-stance, reducing the waste of inputs and increas-ing the productivity of small cultivated areas, at

1This work was partially supported by the UTFORSKPartnership Programme from The Norwegian Centre forInternational Cooperation in Education (SIU), projectnumber UTF-2016-long-term/10097.

2Visiting Research Student at the Norwegian Universityof Life Sciences.

the lowest possible cost (Edan et al., 2009). Cur-rently, there are more than 50 different types ofmobile robots in the world deployed for agricul-tural purposes. These robots, in general, havespecialized accessories, tools and arms to per-form a wide range of agricultural tasks (Becharand Vigneault, 2017; Grimstad and From, 2017).This number gives us an idea of how the poten-tial of robotic technology is embedded in agricul-tural systems. Following this trend, it is possibleto find devices vacuuming apples off the trees inthe US, octopus-like robots for collecting straw-berries in Spain, and machines feeding and milk-ing cows in the UK. These are just a few exam-ples of how robots are taking over fields aroundthe world. These concerns about food produc-tion do not belong exclusively to Europe and canalso be considered in many countries of Central

and South America, particularly, in Brazil. Al-though, the Brazilian agricultural industry has ahigh degree of automation for the planting andharvesting of grains and sugarcane in large ar-eas, farmers still do not use autonomous roboticsystems to perform basic and complex agricul-tural tasks in small areas, such as, vegetable gar-dens and orchards. Some basic agricultural tasksinclude sowing, fertilizing, and irrigation, whilesome complex agricultural tasks consist of har-vesting fruits, killing weeds and plant phenotyp-ing (Bac et al., 2014; Midtiby et al., 2016; Liet al., 2014).

Sugarcane has emerged as an important al-ternative to fossil fuels since, besides to producesugar, it can be used as a clean, renewable sourceof energy reducing the petroleum use - and con-sequently greenhouse gas emissions - and also togenerate other useful products such as, bioelec-tricity and bioplastics (Davis et al., 2009). Sugar-cane farms usually need gigantic areas for plant-ing, which requires many employees and machin-ery, as well as high costs for maintenance and lo-gistics. Moreover, most of sugarcane tasks arelaborious and tedious, susceptible to human er-rors and performed in a harsh environment sub-ject to bad weather conditions, dense vegetation,and wild life. Some of the relevant tasks include:cataloging of plants, identification of weeds/pestsand misplanting (empty spots), classification ofsoil types, and evaluation of plant health.

Figure 1: Conceptual design of the TIBA robot.

Following this motivation, in this proposal weaim to develop the conceptual design and buildthe prototype of a mobile robot to perform au-tonomous tasks of mapping difficult-to-access ar-eas and collecting soil samples in sugarcane farms.TIBA - Tankette for Intelligent BioEnergy Agri-culture - is an alternative solution to deployingseveral robot units in a swarm-based approach tocover a wide area, reducing or eliminating the em-ployment of man-power under unhealthy environ-ments and enhancing the farm logistics. To thebest our knowledge, TIBA (Fig. 1) is the first pro-totype of a 4-wheeled mobile robot to be devel-oped to carry out precision agricultural tasks onsugarcane plantations. Similar systems have beentested in vineyards, orchards and corn productionin the US (e.g., Rowbot).

2 Robot for Sugarcane Fields

In this section, we describe the main aspects of theproposed conceptual design of the TIBA robot interms of mechanical structure, locomotion system,embedded electronics, and software architecture.This design was conceived based on the main chal-lenges found in sugarcane fields, such as tight anddense plant rows, soil mostly composed of sand(66.7%) and clay/mud (33.3%), and possibility ofplant leaves getting trapped in the robot parts.

2.1 Mechanical Design

The mechanical structure was designed to providea very compact solution yet capable to yield 3degrees-of-freedom, translation x, y and the rota-tion θ. The robot is a 4-wheeled small tank com-posed of a 3mm thick steel chassis and 1.5mmaluminum walls (Fig. 1). There is also an alu-minum lid on the top, in order to cover the robotinterior, protect it against water and dust - whichis also provided by rubber tapes along the lidperimeter - and accommodate sensors or other de-vices over its surface, such as laser scanners, so-lar panels, etc. The robot carcass has also cus-tomized holes to provide mechanical accommoda-tion and view of sensors and cameras to the out-side. The robot interior accommodates the me-chanical arrangement—composed of two motors,two gearboxes, belts/chains, pulleys/sprockets,wheel bearings and adapters—and the support forthe electronics system, which is composed of alu-minum profiles in such an arrangement that doesnot interfere with the mechanical components.

The main motivation for this project was thereduction of costs. Thus, a driving configurationthat provides the best mobility with the lowestnumber of motors and other mechanical parts hadto be chosen. The skid-steering configuration, asshown in Fig. 2(a), was selected as a suitable andlow-cost solution, which only demands the em-ployment of two motors, being the curved move-ments achieved by applying differential speeds ineach side. In addition, the trajectory control ofskid-steering robots is already established in theliterature, and its implementation can be based onseveral recent works, such as (Tchon et al., 2015).In TIBA, one motor is designated to drive the twowheels of the same side, as shown in Fig. 2(b).Those wheels are connected by a belt, which pro-vides the same rotation rate to them. Finally, onegearbox connect to each motor provides the torqueincrease and speed reduction. A right-angle gear-box model was employed as a solution for the con-fined space.

Aiming to sustain the cost reduction strategy,the manufacturer and models of the motors, gear-boxes and wheels were selected from the set em-ployed in the existing robot Thorvald II (Grimstad

Figure 2: Driving configuration - (a) skid-steering;(b) TIBA mechanical arrangement.

and From, 2017), so as to get out of their al-ready verified advantages, which include reliabil-ity, low-cost, robustness, easiness for prototyping,and compatibility with other components. Thewheels are a customized set of tire and rim suit-able for snow, sand and clay terrain. To define thesuitable set of motors and gearboxes, a calculationof the needed torque had to be performed. Giventhe selected wheel model - with rw =20 cm tire ra-dius - and the required operational terrain - clayand sand - we could estimate the necessary linearforce to overcome the static friction between thetire rubber and the ground.

By considering an estimated robot weight ofM ≈ 130 kg, the Coulomb friction coefficient andthe rolling resistance coefficient between an off-road tire and sand of µs = 0.6 and Crrs = 0.2respectively, and the same coefficients for wetearth road or clay of µc = 0.55 and Crrc = 0.08(Wong, 2008), the necessary force to overcomethe friction applied in each wheel is given byFw = Nw µs, where Nw is the normal force oneach wheel, and µs is the taken as the worst caseamong the terrain possibilities. By consideringg ≈ 9.8m/s2, the normal force is approximatelygiven byNw =Mg/4 = 130×9.8/4 = 318.5N , andhence, Fw = 318.5 × 0.6 = 191.1N . The necessarytorque τw on each wheel is given by τw = Fw rw,which yields τw =191.1× 0.2=38.22N.m. Finallysince each motor drives two wheels at the sametime and considering that this torque is equallydistributed between them along the belt, the nec-essary torque to drive each wheel/belt set is givenby τb = 76.44N.m.

Among the available motor options from theselected manufacturer, the model 3MeN R© BL840provides the higher nominal torque of τm =16 kgf.cm ≈ 1.57N.m. Given that, among theavailable gearbox options from the selected man-ufacturer, the model APEX R© AER070-050 pro-vides a reduction ratio of 50:1. This drives thewheel/belt set with an increased torque of τb =50 × τm = 78.50N.m, which is more than theneeded value τb, as calculated previously.

Without any load, this motor/gearbox com-bination provides an output angular speed of3000 rpm in the input of each gearbox set, andan output angular speed of 3000/50 = 60 rpm ineach wheel, which provides a total linear speed ofv=2π×60×0.2/60 ≈ 1.26m/s or v ≈ 4.52 km/h.

Considering the given environment and opera-tional requirements for the robot, this is a fairlyfast and safe maximum speed.

2.2 Embedded Electronics Design

The embedded electronics architecture of therobot was also designed so as to comply withthe cost and environment requirements, and thusseveral electronic parts from the robot Thor-vald II were used, such as the motor, powerdriver, mechanical relays, computer, rugged ca-ble/connectors, and DIN rail (Fig. 3).

Figure 3: Embedded electronics parts - (a) Motor;(b) NUC computer; (c) DIN rail arrangements;(d) battery pack; (e) joystick; (f) motor driver.

The power is supplied by one 48VDC NiMHBattery Pack, and distributed into two powerbuses of 48VDC and 12VDC. The former feeds thepower controller driver, and the latter supplies thecomputer, peripheral sensors, and a Vehicle Sup-port System (VSS) - which is represented in thefirst version by an Arduino board. The computerIntel R© NUC NUC6I5SYK Skylake gives a com-pact, lightweight and practical solution for a firstprototype, and counts with a fairly powerful pro-cessor Intel Core i5 and several USB ports. Thedriver Roboteq R© FBL2360 is also a compact so-lution with two channels, which allow the controlof two different motors in the same device. Thecontrol signals are communicated from the PC tothe driver via Controller Area Network (CAN),which is an established and robust network forsensors and actuators communication with stricterror detection. A set of DC-DC converters, low-power relays and a Gigabit Ethernet switch givesthe functionality of adding/removing any periph-eral device or sensor from the system, communi-cate them to the computer, and turn them on/offindividually to save energy.

The VSS has the role of this relay activation,reading voltage and current measurements, andcommunicate them to the PC via CAN. The VSSis also integrated with sensors for the electronicssystem self-monitoring, such as the internal tem-perature and humidity. The system also counts atouchscreen panel mounted on back of the robot,

and an XBox wireless joystick for the control inmanual mode. The panel and the joystick antennaare both connected to the PC directly via USB,being the cables routed outside the robot througha waterproof connector. All the electronic deviceswere mounted over a robust and firm aluminumstructure. The DIN-rail was a robust solution forwiring and keeps the connections firm under vi-bration and impact conditions.

Figure 4: Diagram of ROS nodes.

2.3 Software Architecture

The robot software is developed on the modularmulti-threaded control framework ROS (Quigleyet al., 2009), which presents a large communitywith continuous expansion of software modules(Fig. 4). The prototype is teleoperated using aXbox One wireless joystick. The teleop box nodeis responsible to map the joystick axis and buttonsto the desired velocities and gains. The velocitiescan be modeled as an increment of velocities onthe xy axis. Then, a twist message is publish on atopic. The twist converter node subscribed to theTwist topic and calculates the real velocities val-ues. Then, it publishes them on a topic that thesaga base drive chain node is subscribed. Finally,the latter is responsible to communicate with thedrivers via the CANopen protocol.

Figure 5: Assembly of early prototype - (a) cut-ting of steel and aluminum plates by using plasmacutter; (b) welded parts; (c) chassis and carcass;(d) assembly of mechanics system.

2.4 First Robot Prototype

The first prototype, shown in Figures 5 and 6,was constructed in order to validate all the pro-posed mechanical concepts, feasibility of the im-plemented skid-steering configuration, robot be-haviour in the expected terrains, and robustnessof the embedded electronics system to vibrationand environmental conditions. The robot chassisand carcass were basically a welded assembly ofcut and bent metal sheets. Only for this first ver-sion, some modifications and simplifications wereimplemented due to cost reductions, delays in de-livery of equipment and assembly pieces, and ful-fillment of schedules.

Figure 6: Early prototype - (a) top lid and robotclosed; (b) painted prototype; (c) compact elec-tronics system on the support.

Among the most significant simplifications,the robot mechanical structure, except the toplid, was made entirely of steel, so that the partswelding could be performed faster. The desiredwheel bearing model could not be found on timeand, thus, an alternative one from Volvo was used,which was however heavier, and needed severalmechanical adapters for fixation. The belts andpulleys were replaced by chains and sprockets,since the belts would demand a very large widthfor the given torque/power and, hence, the re-design of the electronics support structure. Fi-nally, due to the lack of available encoders, themotor control and odometry were implementedby only using the motors hall-effect sensors. Thiscondition restricted and hindered the motor po-sition control and high accuracy for an odometrymodel, but it was satisfactory for the first tests.

3 Field Tests

The first field tests were performed at the UMOEBioEnergy sugarcane farm, at Presidente Pru-dente, in the state of Sao Paulo, Brazil. Themain motivations for the farm visit and the robottests were: evaluation of the mechanical conceptand proposed dimensions, better examination andlearning of the environment, and investigation offuture modifications and new features, as shownin Fig. 7a. A few sensors were initially employedin this field test, so that the understanding of theenvironment could be quantified in several datato be later analyzed and used to help the inves-tigation of methods for the robot navigation, asshown in Figures 7b and 7c.

The tests were performed in different rounds,being all the sensor data of each round recorded ina different .bag file in the ROS environment. Forthis test, different chains/sprocket configurationswere tested, as shown in Fig. 7d.

Figure 7: Preparation of robot for field test - (a)robot in the field; (b) sensors attached ad hoc tothe top lid; (c) interface for connecting solar sensorand HT sensor to Arduino; (d) electro-mechanicalarrangement for the test.

3.1 Mechanical Concept Evaluation

The robot presented a fairly satisfactory perfor-mance for remote navigation over the sugarcaneoperational terrain, composed of sand and clay.As expected, its behavior was rough for operationsin concrete and asphalt, because the skid-steeringdriving configuration and the selected wheel typerequire that the tire slides over the terrain, asshown in Fig. 8a. The structure proved to be ro-bust to the uneven terrain, and had a remarkableperformance for unexpected crevices on the soil,as shown in Figures 8b and 8c, which have beenformed due to the accumulation of rain water inthat period.

The skid-steering arrangement showed to bevery convenient to achieve the desired motion

and degrees-of-freedom: forward, backwards, ro-tation around own axis, and curves. The infiltra-tion through the sugarcane tunnels was tested forplants with approximately 1, 2 and 3 meters high(Fig. 9). For the first two cases, the robot wasable to travel along the tunnel without any obsta-cle for its motion. In the 3 meters high case, werethe plants are tall enough to fall back towards thetunnel center and create soft obstacles, there weresome points in which the robot width seemed to bewide for this tightness and got the risk to collidewith the plant stems. For all the cases, it was notobserved the occurrence of plant leaves trappingor getting stuck to the structure or the wheels.

Figure 8: Field test in the UMOE BioEnergyfarm, Presidente Prudente, Brazil - (a) robot en-tering the tunnel; (b) uneven terrain and bigcrevices; (c) robot passing over crevice, the struc-ture is tilted and struggles to overcome the valley.

Figure 9: Field test in UMOE Bioenergy farm,Presidente Prudente SP, Brazil - Tunnels ofplants.

3.2 Thermal Mapping

A FLIR R© One V2 infrared camera in an smart-phone was attached to the robot hood in order tocapture thermal images and videos of the sugar-cane tunnel environment. This was firstly pro-posed as a cheap and efficient solution for therobot navigation, which was already proposedand implemented in many recent works, as by(Fehlman and Hinders, 2009).

Figure 10: Thermal map - (a) sunlight, 1m plants;(b) sunlight, 2m plants; (c) sunlight, 3m plants;(d) late afternoon, cloudy, 2m plants.

The test was applied for rides through sug-arcanes with approximately 1, 2 and 3 meters ofheight, and during the sunlight period - around3 p.m. - and in the late afternoon - around 6p.m. and cloudy. The collected data were greatlysatisfactory. The results for the sunlight period(Fig. 10) show that the thermal mapping createsa path along the ground due to the temperaturedifference between the land and the plants. Ifcompared to the image of a regular HD camera,the thermal path is much more distinguishable,and remains clear even when plants are trappingin front of the camera. For 1 and 2 meters tallplants, the ground is much warmer than the sug-arcane, and a cold kerb is also created at the pathborder, since the base of plantation is even colder.For 3 meters tall plants, the sunlight does notreach the ground strongly. Therefore, the ther-mal map shows that the ground is colder than theplants and no kerb is visible. However, even witha smaller temperature difference, the created pathis strong enough to be distinguished.

The results for a cloudy afternoon, as shownin Fig. 10d, and 2 meters tall plants show thatthe heat pattern held strongly, so that clear paths

were still visible. This was valid for all the threeplant heights. Even though the thermal mappingwere not also performed in rainy conditions, thecollected results were greatly satisfactory, sincenavigation by using a small thermal camera wouldbe a suitable low-cost solution, which is one of therobot concept pillars. In addition to the takenthermal images and videos taken from the forwardview, some media was also collected with the cam-era sightly turned to the sides. The data can beused as an input for image processing and detec-tion of the corridor centerline to be followed by therobot. Several techniques could be employed forthat, such as simple and low-cost ones (e.g., colorthreshold), as shown in Fig. 11, or complex ones(e.g., training artificial neural network for imageclassification).

Figure 11: Example of thermal image processingperformed with the collected data from the fieldtest (the centerline is defined by linear regressionof the margins of warm and orange triangle outputfrom a color threshold filter).

3.3 Odometry and Laser Scanner

A 3D environment of the robot model and motionwas created in ROS by using the robot odometry.In this model, the robot position, orientation, lin-ear and angular speeds are calculated from themotors odometry. Typically, the data from bothencoders and a hall effect sensor are used, but thecurrent prototype still carries only the hall effectsensors due to device delivery delays. The preci-sion obtained by the hall sensor effect was con-sidered satisfactory for the implementation of thefirst 3D model. The goal of the 3D environment isto make possible the simulation of the robot mo-tion, the visualization of sensor collected data -depending on the used platform - and hence theinvestigation of different techniques for the robotnavigation and autonomy.

For this prototype, the data of two laser scan-ners attached to the robot top lid - LIDAR R© A2and Hokuyo R© - were collected and viewed in the3D environment together with the robot model(Fig. 12). Notice in Figure 13 that an interesting

corridor of points was observed with the visualiza-tion of the Hokuyo laser data, which is the scan-ning of the existing plant corridor deployed alongthe sugarcane tunnel. Even though the precisionof those laser scanners may not be the most rec-ommended for this application - due to the highdensity of sugarcanes in such an environment - theobserved data clearly motivates the use of thosedevices to aid in a future robot mapping system.

Figure 12: Laser data and 3D environment map-ping - (a) LIDAR; (b) Hokuyo.

Figure 13: Corridor of points formed due to theexisting corridor of plants (sugarcane tunnel) - (a)perspective view; (b) top view.

3.4 Solar Sensor

One Solar Mems R© NANO-ISSX-60 sun sensor wasattached to the robot top lid in order to detectthe direction of the sunlight (Fig. 14). The sensorpower and signals were connected directly to the

Arduino, and them communicated to the robotcomputer via CAN network. Solar sensors consistof a simple embedded technology that make pos-sible the calculation of the sun ray incident vectorgiven its projection angles along orthogonal axes.The sunlight is detected by a quadrant photode-tector device aligned to the x- and y-axes, fromwhere four voltage levels denoted by V1, V2, V3,V4 are promptly read. The projection angles αx

and αy are hence given by:

αx = tan−1 (C Fx) , αy = tan−1 (C Fy) , (1)

where

Fx =V1 + V2 − V3 − V4V1 + V2 + V3 + V4

, (2)

Fy =V2 + V3 − V1 − V4V1 + V2 + V3 + V4

, (3)

and C is a parameter that depends on the selectedsensor model (SolarMEMS, 2017).

Figure 14: (a) Sensor NANO-ISSX; (b) photode-tector cells; (c) sunlight projections along axes.

The use of solar sensors in the first tests wasinitially motivated by their low-cost, simplicity forinstallation, and some important works that em-ployed them for research with odometry (Lambertet al., 2011), attitude control (Ortega et al., 2010),and navigation (Volpe, 1999). For this robot, themain motivation was to test the device in the sug-arcane environment and track the behaviour of thecollected result in both sunny and cloudy weatherconditions and uneven terrain, which can cause aconsiderable disturbance in the expected results.One of the test goals is to determine the feasibil-ity of using solar sensors to aid the robot initialheading for navigation. Another goal is relatedto a possible future improvement for this robot,which is the use of solar panels for the power sup-ply. The solar sensor may help to control them toan efficient orientation configuration that can getthe most out of the present sunlight.

3.5 Temperature and Humidity Sensor

A temperature and relative air humidity (HT) sen-sor - model RHT03 - was attached to the robottop lid, and electrically connected to the Arduino.HT sensors are strongly used in self-monitoringof embedded electronics systems against overheat

and corrosion due to sea air. Their use was moti-vated for our environment, since it is subjected tohigh temperatures and humidity due to the rain.Future implementations may be motivated by theuse two HT sensors: one within the chassis forthe electronics self-monitoring, and another on therobot surface for the environment monitoring.

4 Concluding Remarks and Perspectives

The designed robot showed to be an efficient low-cost solution for deployment in gigantic sugar-cane fields, with good motivation for operation inswarm-based approach. Since the conceptual testphase of the robot design was concluded and thecollected data was analyzed, some future develop-ments should be implemented.

4.1 Mechanical Modifications

Although the mechanical concept presented agood performance, the wheel fixation to the robotchassis is worryingly stiff, while the chassis by it-self is worryingly flexible. A proposed modifica-tion for the next robot version is to utilize steeringaxis damper along the wheel axes, and a steel barweb crossing the chassis interior in order to en-hance the structure stiffness. The selected wheelbearing model for the first prototype required theuse of several ad-hoc mechanical adapters, whichincreased the robot weight, cost and complexity.

For the next version, a new wheel bearingmodel will be used. The right-angle gearboxshowed to be a good choice for the mechanicalarrangement. Alternatively, we can use a straight-angle gearbox embedded in the wheel, which is amore compact solution. A more suitable and eco-nomic power/motor selection should be performedfor the next version, because this was oversizefor the first prototype, given that the power re-quirement scenario was still unknown. The robotwidth should be reduced so that the robot canfit between the sugarcane tunnels in higher plantscenarios, such as 3 meters tall tunnels. Finally,some original design features should be broughtback, such as: aluminum carcass walls - in or-der to reduce the weight and power demand - andbelts/pulleys in place of chains and sprockets - inorder to reduce noise and backslash.

4.2 Improvement of Embedded Electronics

The embedded electronics system showed to berobust to vibrations and impacts. However, itsmain components were also oversized for the firstprototype in terms of power and processing ca-pacity. Planning the cost reduction for the nextrobot version, the computer should be replacedby a more inexpensive solution, such as a Rasp-berry Pi 3. The VSS system will be composed of acustomized printed circuit board (PCB) based on

ATMEL AVR microcontrollers, which will replacethe Arduino, the current device power connectionsand many wiring. This will better organize thepower distribution and enlarge the possibilities ofcommunication, integration of new sensors, andprogramming. Originally, the idea of using so-lar photovoltaic (PV) panels for the power supplywas being studied. However, for the current de-sign, they became unrealistic, if we consider thetechnology efficiency limitations and the robot di-mensional design constraints.

In future developments close to the availablecommercial version, the robot can have signifi-cantly weight reduction and more efficient powerdrivers. This can bring, together with possibleand notable advances in Solar PV technology, thepossibility of re-including this in the robot designand giving to it the expected power autonomy.

4.3 Software, Control and Navigation

The software architecture should go through sev-eral implementations in order to provide the ex-pected features for the next robot versions, suchas: trajectory control, navigation, communica-tion of computer with the VSS/self-monitoring,and autonomy. The trajectory control is the firststep for the implementation of the robot auton-omy, since it should be able to follow a predefinedpath along the sugarcane tunnel. For the navi-gation through the tunnels, different techniquesbased on the available sensors are being investi-gated, such as: following a corridor of points gen-eration by the laser scanner mapping, following apath defined by the thermal mapping, and usingGPS/IMU to follow a path of known points withpredefined GPS geographic coordinates, which aregiven by UMOE BioEnergy. The combination ofthose methods together with robot mapping tech-niques may boost the definition of an efficient andlow-cost solution for the robot navigation.

4.4 New Features

During the farm visit, some demand for new fea-tures were investigated, being the collection ofsugarcane samples and pest control the most re-markable. The sampling is a frequent procedureperformed in the UMOE BioEnergy fields, whichaims to detect pests in the sugarcane. This is atedious procedure subjected to extreme weather,since the company staff may need to travel a longdistance to the desired sampling point within thefarm area, infiltrate the sugarcane tunnels andslice lengthwise or crosswise one piece of the sug-arcane stem (Fig. 15a).

For the next robot version, one possible so-lution for this problem is to attach a low-costlightweight robotic arm on the robot chassis.Since this is a simple and low-payload task, thereis no need of a complex structure with several

Figure 15: Sugarcane sampling in UMOE BioEn-ergy fields - (a) current procedure, the stem shouldbe sliced or cut crosswise; (b) STEM, a compactsolution for a lightweight sampling arm.

degrees-of-freedom. An interesting solution to beinvestigated is a serial manipulator formed bymultiple Storable Tubular Extendible Members(STEM R©, U.S. patents 3,144,215 and 3,434,674),which are measure tapes that extend as prismaticjoints and retracts inside their chambers in a com-pact solution, as shown in Fig. 15b. Few studiesabout this structure can be found in the litera-ture, such as (Seriani and Gallina, 2015). Finally,the pest control is a required feature to overcomethe frequent problem of pest incidence. Someapproaches for pest detection are being investi-gated, such as a drone for capturing aerial imageswith the Normalized Difference Vegetation Index(NDVI) applied. The collected images of the samesugarcane field at UMOE BioEnergy farm showedto be promising to detect the location of pest andsend the robot directly to those points for a preciseapplication of the pest control substance.

References

Bac, C. W., van Henten, E. J., Hemming, J.and Edan, Y. (2014). Harvesting Robotsfor High-value Crops: State-of-the-art Re-view and Challenges Ahead, Journal of FieldRobotics 31(6): 888–911.

Bechar, A. and Vigneault, C. (2017). AgriculturalRobots for Field Operations. Part 2: Oper-ations and Systems, Biosystems Engineering153: 110–128.

Billingsley, J., Visala, A. and Dunn, M. (2008).Robotics in Agriculture and Forestry, inB. Siciliano and O. Khatib (eds), SpringerHandbook of Robotics, Springer Berlin Hei-delberg, pp. 1065–1077.

Davis, R. J., Baillie, C. P. and Schmidt, E. J.(2009). Precision Agriculture Technologies -Relevance and Application to Sugarcane Pro-duction, The 2009 CIGR Int. Symp. of theAustralian Soc. for Eng. in Agriculture.

Edan, Y., Han, S. and Kondo, N. (2009). Au-tomation in Agriculture, in S. Y. Nof (ed.),

Springer Handbook of Automation, Springer-Verlag Berlin Heidelberg, pp. 1095–1128.

Fehlman, W. L. and Hinders, M. K. (2009).Mobile Robot Navigation with Intelligent In-frared Image Interpretation, Springer Science& Business Media.

Grimstad, L. and From, P. J. (2017). ThorvaldII - a Modular and Re-configurable Agricul-tural Robot, Proc. of the 20th IFAC WorldCongress, Toulouse, France, pp. 4588–4593.

Lambert, A., Furgale, P., Barfoot, T. D. and En-right, J. (2011). Visual Odometry Aided bya Sun Sensor and Inclinometer, IEEE Aero.Conf., Big Sky, MT, USA, pp. 1–14.

Li, L., Zhang, Q. and Huang, D. (2014). A Re-view of Imaging Techniques for Plant Pheno-typing, Sensors 14(11): 20078–20111.

Midtiby, H. S., Astrand, B., Jørgensen, O. andJørgensen, R. N. (2016). Upper Limitfor Context–based Crop Classification inRobotic Weeding Applications, BiosystemsEngineering 146: 183–192.

Ortega, P., Lopez-Rodrıguez, G., Ricart, J.,Domınguez, M., Castaner, L. M., Quero,J. M., Tarrida, C. L., Garcıa, J., Reina,M., Gras, A. et al. (2010). A Miniatur-ized Two Axis Sun Sensor for Attitude Con-trol of Nano-Satellites, IEEE Sensors Jour-nal 10(10): 1623–1632.

Quigley, M., Conley, K., Gerkey, B., Faust, J.,Foote, T., Leibs, J., Wheeler, R. and Ng,A. Y. (2009). ROS: an open-source RobotOperating System, Proc. of the ICRA Work-shop on Open Source Software.

Seriani, S. and Gallina, P. (2015). A StorableTubular Extendible Member (stem) ParallelRobot: Modelization and Evaluation, Mech-anism and Machine Theory 90: 95–107.

SolarMEMS (2017). Solar MEMS TechnologiesS.L. - Sun Sensor NANO-ISSX/c - TechnicalSpecifications, 1.04 edn.

Tchon, K., Zadarnowska, K., Arent, K. et al.(2015). Modeling and Control of a Skid-steering Mobile Platform with Coupled SideWheels, Bulletin of the Polish Academy ofSciences Technical Sciences 63(3): 807–818.

Volpe, R. (1999). Mars Rover Navigation Re-sults using Sun Sensor Heading Determina-tion, IEEE/RSJ Int. Conf. on Intell. Rob.and Syste., Vol. 1, pp. 460–467.

Wong, J. Y. (2008). Theory of Ground Vehicles,4th edn, John Wiley & Sons, Inc.