The Modeling, Construction and Test Process of a 3D Printable Smart

Robot Rider

ao Martin Gorte Jr.

, Witenberg Santiago Rodrigues Souza

2 a

, Artur Benevenutto Coelho

and L

ı Exequiel E. L

opez

Department of Electrical Engineering, Higher Education Institute of Brasilia, L2 Sul, Brasilia, Brazil

Postgraduate Program in Mechatronic Systems, University of Brasilia, Brasilia, Brazil

Department of Computer Engineering, Higher Education Institute of Brasilia, Brasilia, Brazil

Keywords:

Robot Design, Humanoid Robot, 3D Printing, Control Development, Automation.

Abstract:

This work aims to present a modeling procedure, manufacturing process, electronic design, programming and

communication system of a cyclist robot. The robot rider has the ability to ride a bicycle independently by

making decisions on shifts in balance that may occur during its motion by correcting its direction, balance

adjust, and stop. Robot decision-making processes were embedded in a shared control system of interactions

with the external environment by commands sent through wired serial link or radio frequency(RF), allowing

arm and leg- movements. This was achieved using a dual board synchronization powered by Atmel micro-

controllers (Arduino) and fed by information from an electronic gyroscope sensor, an accelerometer, and

remote radio frequency (RF) receptor. The ﬁnal prototype was able to pedal from inertia accelerating gradually

to its full speed, stop the movement, and move its arms to recover balance.

1 INTRODUCTION

The possibilities of autonomous bicycles and self-

controlled or shared controlled robots able to maneu-

ver vehicles have been explored for years (Getz and

Marsden, 1995; Buss, 2000). It can be considered

a theme driven by urban mobility needs, reduction of

human physical effort and aiding people with disabili-

ties. Moreover, modern urban mobility demands echo

applications also motivate the green solution for liv-

ing in large cities. Thus, increasing capacity of robots

moving similarly to humans is essential for this pur-

pose.

One of the main goals in Robotics is the automa-

tion of tasks done by men (Pazos, 2013). Thus, one

of the goals in the ﬁeld of Robotics is to perform

simulation of variety of movements, decision mak-

ing, hence human tasks performance. Despite intense

research activity, large scale implementation of such

tasks for daily life routine were not feasible in the

early robotics, due to high cost and early development

of auxiliary technology, causing to the ﬁeld a moder-

ate development outside of industrial applications.

Moreover, development of such robotic systems

https://orcid.org/0000-0002-2208-7958

requires multidisciplinary knowledge. This work in-

volved mathematical modeling, mechanical drawing,

manufacture techniques in 3D printing, electronic de-

sign, micro-controller programming (Arduino, 2015;

Banzi, 2013) as well as access to auxiliary devices in-

cluding gyroscope, accelerometer, and RF system.

1.1 The Need for Autonomous Vehicles

and Free Space

In many urban areas, human effort can be measured

by the time spent on daily transit, from home to work

and vice-versa. Let the distance of 1km be the initial

reference to this example. In numerous cities, the ur-

ban trafﬁc ﬂows under the average speed of 15km/h, a

cyclist can also perform the same speed, and a pedes-

trian, which speed is about 3km/h, would take 20 min-

utes to perform the same 1km task. On a daily ba-

sis, time is a precious resource and should be spent

wisely . However, in rush hours (from 5 p.m. to 7

p.m.), for instance, places such as Brasilia, Brazil, a

car runs at 12km/h only which means that it takes 5

minutes to perform the 1 km distance. Now, at 1 hour

frame, a bicycle, with the above average speed, would

make 15km, while the car, only 12km in Brasilia, re-

Gorte Jr., J., Souza, W., Coelho, A. and López, L.

The Modeling, Construction and Test Process of a 3D Printable Smart Robot Rider.

DOI: 10.5220/0007773901650172

In Proceedings of the 16th International Conference on Informatics in Control, Automation and Robotics (ICINCO 2019), pages 165-172

ISBN: 978-989-758-380-3

165

sulting in a delay of 15 minutes from a car to a cy-

clist who would have performed 15km for the same 1

hour. In addition, in cities such as Sao Paulo, where

the recorded average speed was below 8km/h, a bi-

cycle could outperform a car by 1 hour. The result

is one hour less of physically and mentally effort that

could be avoided when a bicycle is used (Lopes et al.,

2018).

Although bicycles are great candidate for enhanc-

ing urban mobility, the very action of balancing on a

two wheel vehicle is not an easy task, even for some-

one with no physical or mental disability. Depending

on the motor and sensory limitations of a person, the

task of riding a bicycle becomes even harder, if not

impossible. Additionally, a regularly used road lane

can be used safely by 6 bicycles and the space needed

for parking a bicycle takes one ﬁfth of the amount of

space of a car. Furthermore, automobiles cost on av-

erage 10 to 50 times more than bicycles.

The bicycle is invaluable, yet is presents a com-

plex challenge to solve in robotics, speciﬁcally in

combination with a humanoid rider. However, suc-

cess in this project could possibly result in a new way

of transporting people, animals, and goods.

2 RELATED WORK

Over recent decades, the research on bicycle model-

ing resulted in many mathematical models of balance

and trajectory systems (Getz and Marsden, 1995).

The model validation is another task, performed by

analytic process and simulation to select the most rel-

evant model parameters (Buss, 2000). For instance,

(Buss, 2000) suggests torque and energy conservation

as parameters to be part of simulation. It increases

simulation conﬁdence, given that, these parameters

may be used to determine the system’s stability for

both cyclist and bicycle (Getz and Marsden, 1995).

2.1 A Model for Rigid Bodies

During simulation phase, both cyclist and bicycle may

be represented in several means to reduce irrelevant

features. For example, (Getz and Marsden, 1995)

consider mass m and its height from the bicycle, row

angle α, direction angle from x axis, distance b be-

tween wheels, handle bar angle φ and its inertia J,

and a force µ which according to Getz simpliﬁes the

model.

2.2 System Dynamics

Previous projects have often used Lagrange and

Euler-Lagrange methods (Sharma and Umashankar,

2006) to describe movements of a bicycle with

Denavit-Hartenberg algorithm for the joint connec-

tions of the bicycle. According to (Yetkin and et al,

2014) Lagrange’s equation describes energy conser-

vation deﬁned as

L = T −U (1)

where T is the kinematic energy and U is potential

energy of the system. Another research used Newton-

Euler and Kane equations and virtual work princi-

ple (Shin and Lee, 2004). These methods, however,

are applied with simpliﬁcation and restrictions lead-

ing to difﬁculties to reach acceptable results in real

time (Umashankar and Sharma, 2006).

2.3 Proposals without Artiﬁcial

Intelligence

The majority of proposals without artiﬁcial intelli-

gence contributed to the ﬁeld by exploring which

variables are more relevant to the dynamic model

of a bicycle. The robot system used in (Yam-

aguchi, 2013)model has a remote control trajectory

and PID control system (Yamaguchi, 2013)) powered

PRIMER-V2 model (39cm height) by assisting its ini-

tial movement, its balance on the bicycle, and the

brake action. The system required a remote control

to determine its trajectory.

Most research seeks to solve both balance and

trajectory control challenges. Another end-goal is

developing an anti-collision system and a quality

measurement system of the road where the bicy-

cle passed (Stasinopoulos et al., 2015). In some

cases, researchers try to solve both problems of bal-

ance and trajectory only through validation on sim-

ulations (Sharma and Umashankar, 2006), (Bickford

and Davison, 2013), while other groups also built pro-

totypes in different scales.

(Yetkin and Ozguner, 2013) added to their model

and apparatus to simulate the gyroscope principle.

A disc added below the pedals axis on their model

spins to control the bicycle’s row angle, however, their

model does not have a direction bar. The model pro-

posed by (Bickford and Davison, 2013)presents mul-

tiple closed loops. Their work follows the same path

as many others who developed their work based on

Langrange’s method for dynamic equation and con-

sider the bicycle only, without a rider. As in the pre-

vious research projects mentioned above, the linear-

ity in this model presents limitations and depends on

ICINCO 2019 - 16th International Conference on Informatics in Control, Automation and Robotics

166

nearly constant linear velocities, while others could

start from inertia (Keo and Yamakita, 2009).

Some research, in addition to simulation also

built prototypes for real world tests. Among those

projects, several proposals used balance control meth-

ods such as PD or PID control (proportional integral

and derivative) applied to the direction bar, and gy-

roscope principle whether upon the the bicycle or be-

low it together with PD control (Wang et al., 2012).

There were also variable systems with linear parame-

ter (VLP) (Cerone et al., 2010), springs(Pandey et al.,

2015), and more complex systems such as slide con-

trol altogether with gyroscope principle. Those proto-

types were developed, in majority, from commercial

bicycles with further adaptations. However, there are

also those who built their bicycles in different scales

(Aphiratsakun and Techakittiroj, 2013). According to

(Aphiratsakun and Techakittiroj, 2013), controlling a

prototype by a micro-controller is a very difﬁcult task.

2.4 Proposals with Artiﬁcial Intelligence

(AI)

Although ﬁrst models with AI have been developed at

least two decades ago (Randløv and Alstrøm, 1998),

there were more work with traditional controllers. On

the other hand, the few number of works in AI, show

how the solutions are complex among these groups as

IA solutions claimed to be of simpler implementation

compared to traditional controllers. (Cook, 2004)) for

instance, sought demonstrating to be possible maneu-

ver the bicycle balanced along a trajectory using only

two neurons as controllers. Although the artiﬁcial

neural network (ANN) approach seems to be easier,

it has some drawbacks related to the number of iter-

ations it required until the ﬁrst results are achieved.

Moreover, determining weight was very challenging,

thus leading to error most of the time. (Randløv

and Alstrøm, 1998)tried solving those issues combin-

ing reinforcement learning and calibration, yet their

approach required 5700 attempts for reasonable re-

sults. Without calibration, however, it would be nec-

essary 10

attempts. (Cam et al., 2013) revisited

Randløv and Alstrøm work on Python environment

Panda3D (Goslin and Mine, 2004) and PyBrain sup-

port (Schaul et al., 2010). Their results had shown

that the learning algorithm continuously dropped pre-

vious steps out, making it very sensitive to local min-

ima, which prevents the algorithm’s convergence for

balance and travel tasks. Convergence is the global

learning point for certain problems where reward is

maximized while penalty is minimized.

(Umashankar and Sharma, 2006) presented an in-

teresting method based on neuro-fuzzy adaptive con-

trol where fuzzy rules are applied upon a three layer

neural network. The ﬁrst layer contains inputs, the

second layer has fuzzy rules, and the third layer is

the output. Their model is based on Euler-Lagrange

as the same presented in (Sharma and Umashankar,

2006), however, its new control provides more pos-

sibilities because of its learning capacity altogether

with fuzzy rules showed to be a better option than the

old method which did lose stability after ten seconds.

A complete work in computer graphics focused

on the state description of a cyclist-bike system and

presented a machine learning proposal for a charac-

ter able to perform several stunts on a BMX model

(Tan et al., 2014). In their work, it was not the bi-

cycle that was controlled, but the cyclist by an evolu-

tionary neural network. Due to the stunt’s complexity,

one cannot take into consideration bike dynamic only.

Rather, cyclist and bicycle as two rigid bodies con-

nected with six degrees of freedom using Jacobean

transpose matrix. The project also simulated power

transfer from feet to pedals, presenting excellent re-

sults on stunts performance during simulation. Other

research applied Deep Deterministic Policy Gradient

(DDPG) which is a deep learning approach to con-

trol balance and bicycle’s trajectory by minimizing

the system error (Chung et al., 2017).

3 METHODOLOGY

The initial phase was the creation of mathematical

models for arms and legs. The ﬁnal step involved the

simulation of movements for arms and legs as well

as the ﬁrst real world test for a pedal movement and

balance. The construction of the robot was designed

upon the use of renewable materials such as plas-

tic which reduced the manufacturing cost. The same

plastic was also used to built the 3D printer which has

refurbished pass-motors. Moreover, discarded pieces

of aluminum compose the frame of the bicycle and

the side shields on the legs of the robot.

3.1 Mathematical Modeling

The model followed the Homogeneous Transforma-

tion method. According to Figure 1, the robot has

three actuators (shoulder, arm, and elbow) for each

arm. Using the Kinematic diagram from Figure 1, the

relation from shoulder to hand can be expressed by

the homogeneous matrix equation in eq. 2 composed

by the rotation matrix and the displacement vector.

H(arm)

= H(shoulder)

H(arm)

H(elbow)

(2)

The Modeling, Construction and Test Process of a 3D Printable Smart Robot Rider

167

Figure 1: Arm kinematic diagram: a) link lengths ans rota-

tion angles, b) rotation frames.

Similarly, the mathematical model for the robot

leg was obtained from the kinematic diagram in Fig-

ure 2. The ﬁnal homogeneous matrix expressed by

eq. 3 is the hip-to-foot relation.

H(leg)

= H(hips)

H(thigh)

H(knee)

H(ankle)

(3)

Figure 2: Leg kinematic diagram, a) link lengths and rota-

tion angles, b) rotation frames.

3.2 Mechanical Structure: Hardware

3.2.1 Modeling Each 3D Part

AutoCAD was used as the main graphic tool to design

the robot and bicycle. In the ﬁrst stage, each part of

the robot was sketched, as seen in Figure 3,and later

sculpted to three dimensions. In the second stage, ev-

ery part was converted to STL ﬁle format to be printed

in 3D, Figure 4. Lately, the slicing were made on a

Reprap 3D printer software in order to set how many

layer of PLA would be used and also for time estima-

tion to construct each part. During construction stage,

tests had been performed in order to probe mechanical

properties of critical parts.

Figure 3: Robot (a) and its bicycle(b) both sketched in Au-

toCAD, (c) 3D model.

3.2.2 The Bicycle Construction

The acrylic wheels of the bicycle were estimated ac-

cording to the robot legs with 19cm diameter, 10mm

in thickness, and central hole of 1mm diameter. The

tire of each wheel was made of EVA tape(2.5 mm

ICINCO 2019 - 16th International Conference on Informatics in Control, Automation and Robotics

168

Figure 4: Parts in the printing process servo’s socket on the

left and the slicing process on the right.

thick). This tape has the purpose to increase grip be-

tween the wheel and the ground, and it also reduces

vibration on the composition. The main structure of

the bicycle is made of square bars of aluminum 6cm

thick, blend by aluminum soldering. A tripod was

manufactured to support the bicycle balanced dur-

ing tests. The transmission systems was built with a

sew machine rubber chain, connecting the 3D printed

gears of the bicycle. Pedals were challenging, for it

was made by using spin socket of an orthodontic ma-

chine had been couple on an 1.5mm diameter axis to

smooth movement and reduce grip.

3.2.3 The Power Supply

We used a speciﬁc battery of a hospital appara-

tus (pulmonary apparatus) to provide adequate cur-

rents and voltages for the entire circuit and motors.

This battery contains a circuit with fuses of protec-

tion against surges, under voltages and thermal fuses

for monitoring the temperature. We also manufac-

tured a board able to perform pulse width modulation

(PWM). Is was composed of a voltage regulator of

11V to 5V and a current limiter of 7A to 3A to supply

suitable for the robot. The board contains four 2n3055

voltage regulators and Zener diode for 6V voltage ref-

erence.

3.3 The Automated System Proposal

The algorithm was developed in C++ and deployed

in Arduino board Nano 3.0 with supporting libraries

servo.h and IR.h for servos’ control (Cormen and

et al, 2002; Sedgewick, 1998). Figure 5 show the

ﬂowchart for the algorithm developed. The data from

gyroscope collected through serial protocol RS232

was used to calibrate the end angles of motion.

3.3.1 Detailed Movement

During the tests, the robot rider was able to perform

basic movements, such as pedaling, holding its feet

on the ground while riding, and moving its arms ac-

cording to the feedback from the accelerometer. We

divided the robot’s movements into stop mode and

motion. When it is in stop mode, the robot holds

its feet on the ground. In the transition from stop to

motion, one foot is set to back position prior to ped-

aling, which means the other foot remains the right

position to impulse the robot and overcome the iner-

tia. In intial tests, the arms stayed stretch horizontally

during real world tests a T-shape with the body. The

acceleration of the robot either increase speed or slow

down helping synchronization of both legs during the

required movements. When the robot starts the tran-

sition from movement to complete stop, it decelerates

and takes both legs off of the pedals stretching them

to touch the ground. Both movements, pedaling and

stopping can be remotely set.

Another featured movement built in the algorithm

is the emergency stop. When the gyroscope presents

skewed data, it triggers the movement, which indi-

cates that the robot may be falling. Thus, when the al-

gorithm detects this situation, the robot halts its pedal

function to enter in emergency mode, stretching the

arms and the legs.

Figure 5: Algorithm’s ﬂowchart demonstrating the decision

steps the robot may take.

4 RESULTS

4.1 The 3-D Printing Phase

4.1.1 Process of 3D Printing

The majority of the parts of the robot (approximately

70%) were made by 3D printing. The 3D print-

ers were also assembled using reusable materials. It

The Modeling, Construction and Test Process of a 3D Printable Smart Robot Rider

169

works with plastic wires (3 mm or 1.75mm) made of

polylactic acid (PLA)which is melted to make the lay-

ers of the 3D printed parts.

After every part was ﬁnished by the printer, it was

measured in order to check if its dimensions would

met the designed project. Symmetry, and layer size

were measured. Some parts which should be coupled

to servomotor had density increased by 90% in order

to bear forces applied to them. Other parts were re-

designed for their size were not feasible to be printed

as a single piece or they had inner connections that

would not able accessed such as chest, battery case,

and head). Because of lack of precision for small

parts, some parts had to be adapted, such as the hands.

Figure 6: Printer and the ﬁrst prototype printed legs.

4.2 The Assembling Process

A total 15 servo motors of 13N each with were used to

move all members of the robot. This phase involved

40% of the project budget. The servo motor were also

disassembled to add a 3mm steel axis to the opposite

side the servo axis.

4.2.1 The Bicycle Assembling

In the process of construction of the bicycle, we used

a 5mm solid aluminum bar to let the robot lighter and

balance the weight / power ratio. TIG welding was

used to reinforce the frame and with center of gravity

for the suitable robot. The wheels are made of acrylic

of 10 mm which were cut using the laser process to

stay closer to the CAD model.

After the assembling phase was complete, the pro-

totype, as well as its remote control, as shown in Fig-

ure 7, were ready for testing.

4.3 Simulation and Tests

Legs were tested to check their movement while ped-

aling. Tests had shown a balanced structure and

well synchronized servomotors. After ﬁrst tests, cur-

rent measurements had shown consumption of 1.2A

for each leg, in which every servomotor consumed

400mA. On a standby state the robot consumed

370mA.

Table 1: Time spent on manufacturing each part of the bike

rider.

QUANTITY COMPONENT TIME

01 neck 36min

01 ﬁnger 1 47min

01 ﬁnger 2 47min

02 shin 1h 26min

01 foot 1 1h 49min

01 foot 2 1h 49min

01 neck base 1h 50min

02 thigh socket 2h 25min

02 ebow socket 2h 26min

02 ankle socket 2h 26min

02 forearm socket 2h 26min

02 knee base 2h 42min

02 foot base 2h 46min

02 elbow base 2h 51min

01 waist top 2h 54min

01 waist link 2h 57min

02 knee 3h 4min

02 foot socket 3h 4min

02 thigh 3h 11min

02 knee socket 3h 42min

02 femur socket 4h 4min

01 hand 1 4h 10min

01 hand 2 4h 10min

02 acrylic knee 4h 13min

02 forearm 4h 38m

02 waist socket 4h 43min

02 head 4h 44min

01 hips base 5h 4min

01 battery case 2 5h 7min

01 battery case 1 8h 31min

01 waist 8h 44m

01 hips 8h 48min

01 back 14h 42min

01 chest 22h 1min

51 TOTAL 149h 37min

The ﬁrst pedaling movement had to have ﬁne tun-

ing on tilt coupling, angular speed on each servomotor

and low rotation setup to perform the desired move-

ment.Figure 8 shows that the movements produced by

the system had a squared pattern,requiring additional

ﬁne-tuning.

The algorithm cost required two Arduino boards

(salver and master) to slpit the legs routines from the

signals to the arms. A status signal allows the mas-

ter to know if the other Arduino is on stop mode,

movement, or whether an error occurred. Moreover,

the master board also receives the remote commands

from the joystick. When the system is ready, the mas-

ICINCO 2019 - 16th International Conference on Informatics in Control, Automation and Robotics

170

Figure 7: Assembled robot (second prototype) and joystick

remote control.

Figure 8: Foot movement simulation based on hip-to-foot

displacement vector.

ter Arduino set all initial parameters on the slaver

board to avoid code redundancy. After all parame-

ters set, the slaver board ﬂags back a ready signal to

the master board. The communication between both

Arduinos allows the master to send stimuli signals to

the slaver board when it is in movement mode. When

it is in stop mode however, the master board sends

only one stop signal. If the master board stops work-

ing and the salver board is in movement mode, as it

will not receive the next stimuli, the emergency mode

will be triggered when the time for the next stimulus

is over. On the other hand, if the slaver Arduino board

stops working, the master board will notice it by the

status signal as it stopped receiving from the salver

board. Therefore, the master board will immediately

reset its own status and will keep trying to reestablish

the link with the other board until it succeed. After

reestablishing the link, the master board sends again

the initial parameters to the salver board.

In the simulation phase, the arms was in bent posi-

tion, with hands holding the steering bar as presented

on Figure 9.

Figure 9: Palm of the hand movement expressed by

shoulder-to-hand displacement vector.

5 CONCLUSIONS

This work presented the process of modeling, design,

construction and test of 3D printable robot rider. The

algorithm developed in C++ enable robot to move

its arms and legs to perform desired actions such as

pedaling, turning direction bar, start, and stop move-

ments. The mathematical model proved to reach the

proposed movement, although it required ﬁne tuning

to smooth the movement.

All parts of the bike rider were designed to ﬁt the

servomotors. In order to a better performance, a sec-

ond Arduino board had to be added to split the move-

ments from legs to arms.

Future work will include a free run test with-

out a tripod that supported the initial test of the

robot.Moreover, a forward kinematic system based on

neural network can also be compared to the current al-

gorithm.

ACKNOWLEDGEMENTS

The authors would like to thank the Institutional Pro-

gram for Initiation in Technological Development and

Innovation scholarship (PIBITI-CNPq) and the Scien-

tiﬁc Initiation Program at the Higher Education Insti-

tute of Brasilia- IESB. The authors also would like

to thank Dr. Masahiko Yamaguchi (Dr. Guero) and

Christina Driessen for their wonderful ideas.

The Modeling, Construction and Test Process of a 3D Printable Smart Robot Rider

171

REFERENCES

Aphiratsakun, N. and Techakittiroj, K. (2013). Au-

tonomous au bicycle: Self-balancing and tracking

control(ausub

). In IEEE International Conference on

Robotics and Biomimetics(ROBIO).

Arduino, S. A. (2015). Arduino. Arduino LLC.

Banzi, M. (2013). Primeiros passos com o arduino (First

steps with arduino). Novatec, Sao Paulo, 1st edition.

Bickford, D. and Davison, D. (2013). Systematic multi-

loop control for autonomous bicycle path following.

In 2013 26th IEEE Canadian Conference on Electri-

cal and Computer Engineering (CCECE), pages 1–5.

IEEE.

Buss, S. R. (2000). Accurate and efﬁcient simulation

of rigid-body rotations. Journal of Computational

Physics, 164(2):377–406.

Cam, B., Dembia, C., and Israeli, J. (2013). Reinforcement

learning for bicycle control.

Cerone, V., Andreo, D., Larsson, M., and Regruto, D.

(2010). Stabilization of a riderless bicycle: A linear-

parameter-varying approach. In IEEE Control Syst.

Mag. Citeseer.

Chung, T. et al. (2017). In 2017 14th International Confer-

ence on Ubiquitous Robots and Ambient Intelligence

(URAI), pages 413–417. IEEE.

Cook, M. (2004). It takes two neurons to ride a bicycle.

Demonstration at NIPS, 4.

Cormen, T. and et al (2002). Algoritmos: teoria e pr

atica

(Algorithms: theory and practice). Elsevier, Rio de

Janeiro, 2nd edition.

Getz, N. H. and Marsden, J. E. (1995). Control for an au-

tonomous bicycle. In Proceedings of 1995 IEEE inter-

national conference on robotics and automation, vol-

ume 2, pages 1397–1402. IEEE.

Goslin, M. and Mine, M. R. (2004). The panda3d graphics

engine. Computer, 37(10):112–114.

Keo, L. and Yamakita, M. (2009). Controller design of an

autonomous bicycle with both steering and balancer

controls. In 2009 IEEE Control Applications,(CCA)

& Intelligent Control,(ISIC), pages 1294–1299. IEEE.

Lopes, R. R., Monteiro, L. S. G., de Miranda, R. R., and

de Souza Carvalho, C. (2018). (relat

orio anual de aci-

dentes de tr

ansito - 2017 (annual report on trafﬁc acci-

dents - 2017). Technical report, The Trafﬁc Engineer-

ing Company (Companhia de Engenharia de Tr

afego).

Pandey, A., Mahajan, S., Kosta, A., Yadav, D., Pandey, V.,

Sahay, S., Jha, S., Agarwal, S., Bhise, A., Kumar, R.,

et al. (2015). Low cost autonomous navigation and

control of a mechanically balanced bicycle with dual

locomotion mode. In 2015 IEEE International Trans-

portation Electriﬁcation Conference (ITEC), pages 1–

10. IEEE.

Pazos, F. (2013). Automacao de Sistema e Robotica (Sys-

tem Automation and Robotics). Axcel, Sao Paulo, 1st

edition.

Randløv, J. and Alstrøm, P. (1998). Learning to drive a

bicycle using reinforcement learning and shaping. In

ICML, volume 98, pages 463–471.

Schaul, T., Bayer, J., Wierstra, D., Sun, Y., Felder, M.,

Sehnke, F., R

C1/4ckstie

C, T., and Schmidhuber, J.

(2010). Pybrain. Journal of Machine Learning Re-

search, 11(Feb):743–746.

Sedgewick, R. (1998). Algorithms in C: parts 1-5. Addison

Wesley Longman, New York, 3rd edition.

Sharma, H. D. and Umashankar, N. (2006). A robotic model

(robi) of autonomous bicycle system. In 2006 In-

ternational Conference on Computational Inteligence

for Modelling Control and Automation and Interna-

tional Conference on Intelligent Agents Web Tech-

nologies and International Commerce (CIMCA’06),

pages 107–107. IEEE.

Shin, J.-C. and Lee, C.-W. (2004). Rider’s net moment esti-

mation using control force of motion system for bicy-

cle simulator. Journal of robotic systems, 21(11):597–

607.

Stasinopoulos, S., Zhao, M., and Zhong, Y. (2015). Hu-

man behavior inspired obstacle avoidance & road sur-

face quality detection for autonomous bicycles. In

2015 IEEE International Conference on Robotics and

Biomimetics (ROBIO), pages 2121–2126. IEEE.

Tan, J., Gu, Y., Liu, C. K., and Turk, G. (2014). Learning

bicycle stunts. ACM Transactions on Graphics (TOG),

33(4):50.

Umashankar, N. and Sharma, H. D. (2006). Adaptive neuro-

fuzzy controller for stabilizing autonomous bicycle. In

2006 IEEE International Conference on Robotics and

Biomimetics, pages 1652–1657. IEEE.

Wang, L. X., Eklund, J. M., and Bhalla, V. (2012). Sim-

ulation & road test results on balance and directional

control of an autonomous bicycle. In 2012 25th IEEE

Canadian Conference on Electrical and Computer

Engineering (CCECE), pages 1–5. IEEE.

Yamaguchi, M. (2013). Cyclist robot able to ride

a bicycle (original title in japanese). htt p :

//ai2001.i f de f . jp/primer V 2/primer V 2.html.

Accessed on January 14, 2019.

Yetkin, H. and et al (2014). Gyroscopic stabilization of an

unmanned bicycle. In American Control Conference.

Yetkin, H. and Ozguner, U. (2013). Stabilizing control of an

autonomous bicycle. In Control Conference (ASCC).

ICINCO 2019 - 16th International Conference on Informatics in Control, Automation and Robotics

172