2016 International Electronics Symposium (IES)

End-to-End Coconut Harvesting Robot

Teguh Satrio Wibowo∗ , Indra Adji Sulistijono† and Anhar Risnumawan‡

Mechatronics Engineering Division
† Graduate School of Engineering Technology
Politeknik Elektronika Negeri Surabaya (PENS)
Kampus PENS, Jalan Raya ITS Sukolilo, Surabaya, 60111 Indonesia
Tel: +62-31-594 7280; Fax: +62-31-594 6114
Email: ∗ teguhsatrio@[Link], {† indra, ‡ anhar}@[Link]

Abstract—World coconuts production increasingly demands a

huge number of raw coconuts to be cultivated. This is starting
as everyday needs, foods, traditional cuisines, to many industries
which depend on raw coconuts. While coconut production in
Indonesia has only contributed to 27%, yet Indonesia has the
biggest coconut area in the world. This is because coconuts are
often cultivated conventionally, climbing by men. The condition
is worsen as the number of men climbing coconut trees has
increasingly decreased and industries requires to reduce cost
production. In this paper, we develop an end-to-end autonomous
coconut-harvesting-robot for easy operation and portable to carry
along in large plantations. First, a novel climbing and fastening
mechanism to climb a tree with varying diameter is introduced. Camera
Second, position of a coconut is then detected using steady-state
genetic algorithm followed by sending a command to harvest

the detected coconut using an extendable arm equipped with a    
specifically designed saw. Experiment results show the ability of Wireless
Communication
the robot to harvest coconuts.
Keywords—Coconut harvesting robot, SSGA, steady state ge-
netic algorithm Camera View

Fig. 1. Proposed coconut harvesting robot. The robot is autonomously

I. I NTRODUCTION searching for a coconut reaching the top of the tree harvesting coconut using a
specifically designed saw. User can view its action using mobile phone screen.
World coconuts production is 64,897 million nuts and
Indonesia is one of the leading coconuts production after India
and Philippines. Indonesia contributes to 27% of coconuts
these men usually do not have any insurance coverage, any
production and has the biggest coconuts area in the world.
accidents while climbing the trees would affect the entire
From 2000 to 2011 the area is around 3.69 to 3.91 million
family. In many such households, people suffer a lot because
hectares. Although it is the biggest coconuts area in the world,
this would be the only source of income for their livelihood. In
Indonesia only produces coconuts around 3.04 to 3.25 million
addition, the people who fall during climbing greater heights
tons. As compared to India and Philippines, which have area of
are bedridden for the rest of their life.
2 and 3.8 million hectares respectively and able to contribute
around 90% of coconut production, Indonesia is considered The condition is worsen by several industries, which de-
lower in coconut production which is around 0.8 tons per pend on raw coconuts, for its processing and manufacturing of
hectare. Farmers of marginal income hold roughly 90% of various value-added products. The inflorescence sap of coconut
coconut holding in Indonesia. Moreover, farmers often utilize palms used to produce palm nectar, desiccated coconut and its
coconut trees as everyday needs for which each and every part powder, packed coconut milk, coconut cream, coconut milk
of the tree is used by people. From coconuts and trees what powder, tender coconut water, vinegar, nata-de-coco, etc. are
can be used to produce copra, oil, milk, chips, milk powder, few of them. The growing market for all coconut products
toddy, refined sugar, palm sugar, flower syrup, jam, vinegar, is increasing especially in Indonesia that often uses coconut
jelly, tender coconut water, fuel for traditional cooking, etc. in their cuisine. Hence, any impact in coconut cultivation
Maximizing the abundant amount of coconuts production is sector will adversely affect these industries and their markets.
therefore extremely important in Indonesia. Thousands of people get employment from these industries.
One of the main threats of this industry is the increasing cost
Coconut harvesting is traditionally carried out by men who of production and the decrease in the labour resource.
belong to the economically lower class of society. As the
number of literacy rate increases and awareness about the high In such situation there is genuine need to develop an
paid job opportunities, the number of men climbing coconut autonomous coconut harvesting robot to support the coconut
trees for harvesting coconut has decreased increasingly. As industries. Previous works [1]–[5] on coconut harvesting robot

978-1-5090-1640-2/16/$31.00 ©2016 IEEE 444

only focus on a certain part from the overall system, such as A limitation of their system was that they could not detect
only either climbing mechanism or coconut detection using a partially occluded grapes (crops). However, as they were
computer vision technique, neglecting the fact that the whole performing yield estimation and not accurate crop detection
system is highly important for successful harvesting in chal- they were able to cope with this limitation.
lenging conditions. In this paper, an end-to-end autonomous
coconut harvesting robot system starting from mechanical The problem of apple detection has been explored by Wang
climbing and harvesting designs to vision-based coconut de- et al. [10] so that they could perform yield prediction. The
tection and harvesting is introduced, as shown in Fig. 1. The method was based on color and distinctive specular reflection
robot is portable to carry along in large plantations, easy pattern. Further information, such as the the average size of
to operate, and equipped with a novel climbing and auto- apples, was used to either remove erroneous detections or
fastening mechanism to climb a tree with varying diameter. to split regions that could contain multiple apples. Another
The robot has hexagonal main body, two degrees of freedom heuristic employed was to accept as detections only those
arm, grinstone and camera on top of the arm, three wheels regions which were mostly round.
and one spring in every wheel. The position of a coconut Segmentation approach for sweet peppers has been pro-
is then detected using steady-state genetic algorithm followed posed by Bac et al. [11] and Hung et al. [6] utilized conditional
by sending a command to harvest the detected coconut using random fields (CRFs) for almond segmentation. Bac et al.
extendable arm equipped with a specifically designed saw. aimed to develop a robotic harvesting system and introduced
The rest of this paper is organized as follows. Section II a class segmentation approach in order to build an accurate
describes related work. In Section III explains the methodology obstacle map. They used a six band multi-spectral camera
and it consists of III-A the robot mechanism and control and (with bandwidths of between 40 − 60nm2 ) and used a range
III-B describes the coconut detection algorithm. IV and V of features including the raw multi-spectral data, normalized
describe experiments and conclusion, respectively. difference indeces, as well as entropy based texture features.
Experiments in a highly controlled glasshouse environment
showed that this approach produced reasonably accurate seg-
II. R ELATED W ORK mentation results, however, the authors noted that it was not
accurate enough to build a reliable obstacle map.
Conventional harvesting method on reaching the top of a
coconut tree, to test its maturity the climber taps the nut in On the other hand, Hung et al. developed an almond
the lowermost bunch with its harvesting knife. If the coconut segmentation approach in order to perform yield estimation.
is mature, the bunch at the base of the stalk is cut. When it They proposed a class segmentation approach which learnt
drops to the ground, the climber also cleans the crown and features using a sparse auto-encoder. These features were then
removes the dry leaves, sheaths and spathes. For small trees, used within a CRF framework and was shown to outperform
cutting the coconut bunches with a knife, attached to a long previous work. They achieved impressive segmentation perfor-
bamboo pole will do. Climbing the coconut palm in order to mance, but did not perform object detection. Furthermore, they
stay beneath the leaf crown to harvest fruits and cut old leaves noted that occlusion presented a major challenge.
seems quite simple. But it proves to be much more challenging
to climb up the coconut leaf crown in order to reach the young More recently, tomato detection has been proposed by
inflorescences for making controlled pollinations with bagging Yamamoto et al. [9] by first performing class segmentation.
for research purposes, and for harvesting the toddy (the sweet Color and shape features were used to train a classifier and
sap from coconut inflorescences), which serves to produce regression trees classifier. This produced a segmentation map
sugar, vinegar, wine and spirits. This motivated us to propose and grouped connected pixels into regions. Each region was
robotic harvesting method, where it can peep and reach out to declared to be a detection and to reduce the number of false
any remote nook and corner at the top of the tree. alarms they trained a non-fruit classifier using a random forest.

There is a high risk associated with coconut tree climbers The aforementioned works only focus on a certain part
because accidents can lead to severe physical damage, some mostly on fruits detection neglecting the fact that robot mech-
of which may even be fatal. On the other side there is anism is also important for successful harvesting in challenging
significant increase in price of coconut and coconut products. conditions. Several approaches have addressed only the crop
To tackle all these problems, the process of coconut harvesting segmentation task [6], [11] and not detection. Most of the
automaticallyd reduce the need of human labour and also work that has examined crop detection has predominantly
accident involved in the process. The robotic arm plays a been developed for yield estimation [7], [10] and so accu-
crucial role in making the robotic coconut tree climber a rate detection was not necessary. The limited work that has
successful machine. A robust and proper robotic arm would examined accurate crop detection has done so for crops in
also help in balancing the body of such a robotic climber. controlled glasshouse environments [9]. As such the issue
of crop detection in highly challenging conditions remains
Previous works on robotic crop harvesting, for segmenting unsolved.
and detecting crops, have been explored by several researchers
[2], [6]–[11]. The grape detector of Nuske et al. [7], [8] III. M ETHODOLOGY
was initially proposed and one of the earliest crop detection
systems. A radial symmetry transform was used to detect We design the model using CAD software for easy assem-
grapes in an image and then used this information to perform bly. Hollow aluminium material is mostly used in our design
accurate yield estimation. Human head tracking was proposed because of light weight, availability, low cost, and its strength.
by [12], but the problem is different from coconut detection. The prototype is then made as per CAD design and tested in

445
 3. This suspension mechanism is important as to provide better
stability of the viewing camera for later processing of coconut
detection. The spring which is mounted on the rod parallel
to the pipe and slider, operated by pushing the slider to the
inside so as to sufficiently grab the trunk and adjust to varying
tree trunk diameter. The length of the spring for the maximum
pressure is 5 cm, whereas under unpressed conditions is 15
cm, this is because of the limited expansion length of a slider.
Slider maintains the position of the wheels to be remain
perpendicular to the supporting side of the robot body and
also as a way of translational movement mechanism when
the wheel adjust to the tree diameter. Slider is designed with
stainless steel material and has expansion capabilities up to 10
cm so that when all of the sliders fully suppressed, tree trunks
diameter can reach up to 40 cm. Slider is mounted on the side
of the robot body with connections made of aluminum shaped
elbows using bolts. Fig. 4 shows final assembly of the real
Fig. 2. CAD final assembly of coconut harvesting robot. Note that it has robot.
three auto-fastening sliders to adjust with varying tree diameter.

the real conditions and suitable modification could be made to

the final model.

A. Hardware
The proposed coconut harvesting robot comprises of main
body and robotic arm, as shown in Fig. 2. The body shape of
a hexagon made of hollow aluminium material with a width
of 2.5 x 2.5 cm and a length of 38 cm per side. For portability
and ease of use, the robot is equipped with a clip lock on
one side of the robot frame that serves as a manual lock and
can be opened or closed to be inserted onto the tree trunk.
To deal with varying tree trunk diameter, the robot has an
auto-fastening mechanism using three wheels connected with Fig. 4. Final assembly realization of coconut harvesting robot.
springs on all three sides so that the robot can provide better
grip wrapped around a tree trunk. 2) Robot arm: The arm is made of aluminium with a length
1) Robot main body: The wheel dimension is 8.5cm of of 170 cm. The arm can move up and down using DC motor by
diameter, 21cm wide, and made of durable solid sponge. Se- a pulley. The DC motor rotation is converted into linear motion
lection of this sponge as the wheel surface is highly important using a timing-belt. The arm linear movement is designed to
as it has the characteristic to fully attach on the tree surface. be able to reach the coconuts. At the top arm, a specifically
Each wheel is mounted on the robot base frame with 60 degree designed saw and wireless camera are attached. The saw is
orientation facing to the other wheels to form an equilateral designed using brushless DC motors for high speed rotation
triangle when viewed from above. in order to quickly cut the coconut stem. The blade is a sharp
grinding for cutting.

 Accelerometer Sensor

Central
 Processing

Driver Motor

 

IP Camera

Fig. 5. Electronic circuitry of the robot.

Fig. 3. Auto-fastening mechanism to grab varying tree trunk diameter. (1)
Spring, (2) Slider, and (3) Durable sponge-wrapped wheel. 3) Electronics control: The main control of the robot as
shown in Fig. 5 is using the Arduino Mega 2560. While for
Each wheel has an auto-fastening mechanism using a image processing side a sophisticated Raspberry Pi 2 is used
spring which serves as a suspension as well, as shown in Fig. for computation, portable, and OS-based platform containing

446
many open libraries such as the well-known image processing the mating pair. Variations of this crossover exist. Scattered
OpenCV library. crossover where a vector the same size of the chromosome is
created containing zeros and ones. The new child obtains the
Wireless connectivity is done using a bluetooth module hc- genes of the first parent whenever there is a one on the vector
05 as the sender (transmitter) and a bluetooth module hc-06 and gets the genes of the second parent whenever there is a
as a receiver (receiver). Data will be read and recognized as a zero on the vector. Two-point-crossover is another one where
character ASCII (American Standard Code for Information In- two positions for crossover is selected.
terchange). Data received by the bluetooth will be re-processed
to be converted into movement of the robot in combination Mutation operator is basically randomly changing the
with data from the balance sensor. value of a gene. For real valued genes, this change could
be done by selecting a number uniformly from the range of
B. Software possible values. One other possible way is to use a distribution
like Gaussian with mean zero and a given variance where the
1) Genetic Algorithm: The Genetic algorithms (GAs) [13] number generated by this distribution is added to the value of
are stochastic global search method that imitates the evolution the gene. It may be adjusted such that the variance decreases as
of natural biological process. GAs operate on a population the algorithm iterates. In our analysis for Gaussian mutation,
of candidate solutions applying the fundamental of natural the variance is chosen to be 1 and shrinks linearly to zero with
genetics and survival of the best, to converge to a better each generation.
approximations of solution. At each generation, a new set
of approximations is created by the process of selecting At the beginning of the computation a number of individ-
individuals according to their level of fitness in the problem uals (the population) are randomly initialized. The objective
domain and breeding them together using operators borrowed function is evaluated for these individuals. The first / initial
from natural genetics. Thus, These iterations leads to the generation is produced. If the optimization criteria does not
evolution of populations of individuals that are better suited to meet the creation of a new generation starts. Individuals
their environment than the individuals that they were created are selected according to their fitness for the production of
from, just as in natural adaptation. Due to the GAs extensive offspring. Parents are recombined to produce offspring. All
emulation of nature, biological terms have been borrowed and offspring will be mutated with a certain probability. The fitness
deserve cursory explanation. Accordingly, they are contrasted of the offspring is then computed. The offspring are inserted
by demonstrating their role both in nature, as well as the into the population to replace the parents, produce a new
algorithm. generation. This cycle is performed until the optimization
criteria are reached.
Each individual is called chromosomes. Each basic unit of
a chromosome is called a gene. A chromosome can be thought 2) Steady State Genetic Algorithm: The steady-state ge-
of as a string or a vector where each of its components is netic algorithm [14] uses overlapping populations. In each
a symbol from a possible set of values. The set of vectors generation, a portion of the population is replaced by the
collectively make up the population. The main idea lying newly generated individuals. At one extreme, only one or
behind the GA is the reproduction of a population. Because two individuals may be replaced each generation (close to
of the reproduction, a new population is formed. Although 100% overlap). At the other extreme, the steady-state algorithm
many variations are possible and available, the main parts becomes a simple genetic algorithm when the entire population
of reproduction, analogous to biological reproduction, are is replaced (0% overlap).
selection, crossover, and mutation. Since the algorithm only replaces a portion of the pop-
Selection is an essential part of the natural evolution ulation of each generation, the best individuals are more
process and hence is an essential part of GA. The most com- likely to be selected and the population quickly converges
monly known selection element is roulette-wheel selection. In to a single individual. As a result, the steady-state algorithm
roulette-wheel method, the candidate individuals are given a often converges prematurely to a suboptimal solution. Once
slice of the wheel proportional to their assigned probabilities again, the crossover and mutation operators are key to the
obtained by scaling. Then the wheel is rotated N times where algorithm performance; a crossover operator that generates
N is the number of individuals or parents needed for repro- children unlike their parents and/or a high mutation rate can
duction. Thus, the best individuals have a higher probability delay the convergence.
of being chosen. However, undesired solutions also have a Fitness function is computed using the information from
small chance of survival. More computation time effective way circle Hough transform [15]. Unlike the linear Hough trans-
to select individuals from the wheel is to create one random form [16], the circle Hough transform relies on circle equa-
number and from that point continue in increments of equal tions. The circle equation is,
size (1/N ). Another alternative for selection is tournament
in which chromosomes are selected randomly and the best is
chosen among the them. This procedure is again repeated N r2 = (x − a)2 + (y − b)2 (1)
times. There can also be variations of this tournament where where a and b represent the center coordinates, and r is the
the best string does not always win but wins with a chance of circle radius. The parametric representation of this circle is,
certain threshold.
Crossover of a single point is one of the most basic one.
The position of the crossover is determined randomly. Then x = a + r ∗ cos(θ)
the determined length of the strings is exchanged between y = b + r ∗ sin(θ) (2)

447
 TABLE II. ROBOT WHEEL ROTATION SPEED
In contrast to a linear Hough transform, a circle Hough
transform relies on three parameters, which requires a larger Wheel diameter 80mm
computation time and memory for storage, increasing the com- PWM RPM RPS cm/s
plexity of extracting information from our image. In practice, 128 102 1.7 42.7
we perform Canny edge detector on HSV color domain and 160 130 2.17 54.43
then we code the circle Hough transform by setting the radius 192 159 2.65 66.57
to a predefined value. For each edge pixel, a circle is drawn 224 188 3.13 78.71
with that point as origin and radius r. An array 3-dimensional 255 215 3.58 90.01
with the first two dimensions representing the coordinates of
the circle is used and the last third specifying the radius. The
values in the array are increased every time a circle is drawn meters, so that the robot maximum speed to climb-up above
with the desired radius over every edge pixel. The array, which the palm tree is within 22 seconds, from table II.
count of how many circles pass through coordinates of each
edge pixel, proceeds to a score to find the highest count. The Accelerometer sensor influences the rotation of the wheel
coordinates of the center of the circles in the images are the motor in order to maintain the robot movement to be balanced
coordinates with the highest count. and not skewed. The experimental result is conducted so as
to roughly balance the robot due to high noise which could
be attributed by limited space of sensors placement. Testing
IV. E XPERIMENTS data on accelerometer can be seen in table III. Data sensor
processing was performed using a common low pass filter
Experiments are conducted in coconut plantation Blitar, in order to remove high frequency noise due to undesired
East Java, Indonesia. We select this place as it is nearby Indian vibration.
Ocean and it has highest coconut production in East Java.
Setup of coconut harvesting robot is shown in Fig. 6. TABLE III. ACCELEROMETER (MMA7361) DATA AGAINT ROBOT
POSITION

Position x y z
Normal 450 300 245
Wheel-1 330
Wheel-2 260 238
Wheel-3 270
Wheel-3-1 330 255
Wheel-1-2 230
Wheel-2-3 270

At trial cut coconuts robot body was hit by coconuts

that weighs about 10kg, but the robot does not fall down
when exposed to loads of coconuts, one of them due to the
mechanical spring system that has worked well for pressing
coconut trees. At the cutting blade mechanical systems, the
selection of blades and motors used influence on whether or not
cut stems coconuts and long- or cutting process. In experiments
using fine grinding blades, the blades can cut stems coconuts,
but the cutting process takes a long time.
Fig. 6. Setup of coconut harvesting robot in coconut plantation Blitar, East
Java, Indonesia. Random initialization of steady state GA is shown in Fig.
8 and coconut detection result is shown in Fig. 9. From these
A bluetooth communication is selected as a wireless observations, it is clear that using color alone in a straight
medium because it is still in range of the data sender and forward manner is likely to provide poor results in many cases
an average peak coconut trees, as shown in table I. Moreover as the coconut and leaf color is almost similar. Thus we attempt
it is easier for hardware setup. to find visual signature that is quite robust using circle Hough
transform, as it could be the case of human perception. Fig.
TABLE I. M AXIMUM DISTANCE OF TRANSCEIVING DATA 7 shows coconut harvesting robot has the ability to perform
Transmitter Receiver Maximum Distance autonomous harvesting.
HC-05 HC-06 30 meter
Android HC-06 25-30 meter
V. C ONCLUSION
Wireless Wireless Receiver 10 meter In this paper, an end-to-end autonomous coconut-
harvesting-robot has been developed to solve the problem of
The robot showed relatively well to move up and down low number of conventional men climbing coconut trees to
on a tree, auto-fastening using the spring has shown quite increase coconut production and to support industries that
well in reducing and adapting to varying coconut tree trunk depend much on raw coconuts. Different from previous works
diameter, so that the robot could climb coconut trees quickly. which only focus on a certain part of coconut harvesting
The average height of a coconut tree is approximately 20 system, we have demonstrated the whole system for successful

448
Fig. 7. Proposed autonomous coconut harvesting robot in action.

ahead,” International Journal of Computational Vision and Robotics,

vol. 3, no. 1-2, pp. 4–34, 2012.
[3] A. Abraham, M. Girish, H. Vitala, and M. Praveen, “Design of harvest-
ing mechanism for advanced remote-controlled coconut harvesting robot
(arch-1),” Indian Journal of Science and Technology, vol. 7, no. 10, p.
1465, 2014.
[4] R. K. Megalingam, T. Pathmakumar, T. Venugopal, G. Maruthiyodan,
and A. Philip, “Dtmf based robotic arm design and control for robotic
Fig. 8. Random initialization of GA. coconut tree climber,” in Computer, Communication and Control (IC4),
2015 International Conference on. IEEE, 2015, pp. 1–6.
[5] R. Fernández, C. Salinas, H. Montes, and J. Sarria, “Multisensory
system for fruit harvesting robots. experimental testing in natural
scenarios and with different kinds of crops,” Sensors, vol. 14, no. 12,
pp. 23 885–23 904, 2014.
[6] C. Hung, J. Nieto, Z. Taylor, J. Underwood, and S. Sukkarieh, “Or-
chard fruit segmentation using multi-spectral feature learning,” in 2013
IEEE/RSJ International Conference on Intelligent Robots and Systems.
IEEE, 2013, pp. 5314–5320.
[7] S. Nuske, S. Achar, T. Bates, S. Narasimhan, and S. Singh, “Yield
estimation in vineyards by visual grape detection,” in 2011 IEEE/RSJ
International Conference on Intelligent Robots and Systems. IEEE,
2011, pp. 2352–2358.
[8] S. Nuske, K. Wilshusen, S. Achar, L. Yoder, S. Narasimhan, and
S. Singh, “Automated visual yield estimation in vineyards,” Journal
of Field Robotics, vol. 31, no. 5, pp. 837–860, 2014.
[9] K. Yamamoto, W. Guo, Y. Yoshioka, and S. Ninomiya, “On plant
detection of intact tomato fruits using image analysis and machine
Fig. 9. Coconut detection result. learning methods,” Sensors, vol. 14, no. 7, pp. 12 191–12 206, 2014.
[10] Q. Wang, S. Nuske, M. Bergerman, and S. Singh, “Automated crop yield
estimation for apple orchards,” in Experimental Robotics. Springer,
harvesting in challenging conditions. The robot is easy to 2013, pp. 745–758.
operate and portable to carry along in large plantations. A [11] C. Bac, J. Hemming, and E. Van Henten, “Robust pixel-based classifi-
novel climbing and fastening mechanism to climb a tree with cation of obstacles for robotic harvesting of sweet-pepper,” Computers
varying diameter has been introduced. The position of a co- and electronics in agriculture, vol. 96, pp. 148–162, 2013.
conut is detected using steady-state genetic algorithm followed [12] I. A. Sulistijono and N. Kubota, “Evolutionary robot vision and particle
swarm optimization for multiple human heads tracking of a partner
by sending a command to harvest the detected coconut using robot,” in 2007 IEEE Congress On Evolutionary Computation. IEEE,
an extendable arm equipped with a saw. The experiments 2007, pp. 1535–1541.
show encouraging results which would be beneficial to increase [13] J. H. Holland, “Adaptation in natural and artificial systems, univ. of
coconut production and to support coconut industries. mich. press,” Ann Arbor, 1975.
[14] D. Whitley and J. Kauth, GENITOR: A different genetic algorithm.
Colorado State University, Department of Computer Science, 1988.
R EFERENCES [15] S. J. K. Pedersen, “Circular hough transform,” Aalborg University,
[1] R. K. Megalingam, R. Venumadhav, A. Pavan, A. Mahadevan, and Vision, Graphics, and Interactive Systems, vol. 123, 2007.
T. Charly, “Kinect based wireless robotic coconut tree climber.” [16] D. H. Ballard, “Generalizing the hough transform to detect arbitrary
[2] K. Kapach, E. Barnea, R. Mairon, Y. Edan, and O. Ben-Shahar, “Com- shapes,” Pattern recognition, vol. 13, no. 2, pp. 111–122, 1981.
puter vision for fruit harvesting robots–state of the art and challenges

449