Systems engineering of agricultural robot design

IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS, VOL. 24, NO. 8, AUGUST 1994

Systems Engineering of Agricultural Robot Design

1259

operate in concert? Is it better for all manipulators to do the same tasks in parallel or is it better for some manipulators to act as Yael Edan and Gaines E. Miles servers? When designing a robot, parameters such as speed and acceleration, must be determined. Regardless of the type of drive (i.e., hydraulic, pneumatic, or electric), increasing acceleration raises the Abstract-The design of an agricultural robot is a complex task since price exponentially. Multiple slow arms might be less expensive than in addition to the many closely related design parameters that must be determined, the design is highly affected by crop parameters which one fast arm. Multiple robots can be advantageous when resource are uncertain and loosely structured. This paper presents a systems and task sharing lead to reduced costs. Dividing the task between engineering method to evaluate the performance of an agricultural robot multiple arms can reduce the complexity of each subtask and thereby by simulating and comparing different types of robots, number of arms, reduce sensor and control requirements. This, in tum, can lower the multiple arm configurations, workspace design and dynamic characteriscontrol system complexity and perhaps reduce -costs and increase tics. Numerical simulation tools were developed to quantify measures of machine performance such as cycle time and percentage of sucxxsful productivity. On the other hand, increased level of sharing among cycles based on an extensive statistical analysis using measured fruit robots requires more complex supervised control. Therefore, the locations and simulated crop parameters. The methodology developed was optimum configuration should be based on a costhenefit ratio and applied to determine design parameters for a robotic melon harvester. Simulation results indicated that the Cartesian robot was faster than determined by an economic analysis [21]. An initial step for such the cylindrical robot for the melon harvesting task. Activating two arms an analysis is to determine the performance for different design in tandem was the fastest configuration evaluated. Simulation provided parameters and this is the aim of this paper. an important tool for evaluating the multitade of design and crop Since there are many closely related design parameters, optimal parameters and for comparing alternatives in a timely manner prior to robot selection and design is an extremely difficult and time conprototype construction. Through systems engineering design parameters and preferable crop conditions were recommended based on which a suming task that is not yet automated [16], [25]. The solution for similar problems in industry has been to simulate robot performance prototype robotic melon harvester has been constructed. using commercial software packages that provide animated, graphical representation of the time-varying solutions [ 151, [28]. Engineers I. INTRODUCTION select robots by evaluating, in simulation, altemative manipulators Robots are perceptive machines that can be programmed to perform integrated into a workcell considering different workcell setups, a variety of agricultural tasks such as transplanting, cultivating, material, tool flow and control strategies [lo], [21]. However, the spraying, trimming and selective harvesting. Despite the tremendous agricultural environment is loosely structured, i.e., there is large amount of robotic applications in industry, very few robots are variability between successive tasks and even between successive operational in agriculture production [26]. On contrary to industrial items even though they are nominally identical. Thus, not all condiapplications which are simple, repetitive, well-defined and known a tions can be predicted and performance must be derived based on a priori, an agricultural robot must deal with an unstructured, uncertain statistical analysis [20]. Simulation of robotic systems in controlled and varying environment which cannot be predetermined. Thus, funagricultural environments has been performed to determine optimal damental technologies must be developed to solve difficult problems configuration of the workcell [17] and robot design parameters [7], as: [27]. However, since most agricultural domains are unstructured, Mobile operation in a three-dimensional continuously changing robotic performance must be evaluated for a variety of crop contrack; ditions. In this paper, a systematic method has been developed Random location of targets (fruits); to evaluate robot performance for an unstructured environment. Variability in fruit size, shape, color, texture and firmness; Based on simulations performed for a multitude of design and Delicate products; crop parameters the best set of design parameters can be derived Variable environmental conditions like illumination (due to and guidelines provided to what kind of crop parameters should clouds, sun direction), leaves occlusion; and be adapted to improve performance of a robot operating in an Hostile environmental conditions like dust, dirt and extreme unstructured environment. temperature and humidity. The objective of this research was to develop a systems engineering An important factor in the overall performance of a robotic system approach to evaluate the performance of an agricultural robot peris selecting the most appropriate manipulator for the specific task forming in a varying environment to determine: 1) the most suitable and defining its motions. When selecting a manipulator, the question set of design parameters, i.e, type of robot (Cartesian, cylindrical, is not only which type to use, Cartesian, cylindrical, spherical or spherical); workspace design; number of arms (one, two, three); articulated, but how many and how should they be configured to multiple arm configuration (serial, tandem, parallel); and actuator speeds; and 2) the preferable crop parameters. General simulation Manuscript received April 10, 1992; revised August 1, 1993 and Otober 5, models were developed. Their applicability was demonstrated by 1993. Thies work was supported by Grants No. US-1254-87 and US-1682-89 from BARD, the United States-Israel Binational Agricultural Research and applying them to design a robotic melon harvester. Melons are fleshy Development Fund. delicate fruits. They grow randomly scattered over the beds either Y. Elan is with the Department of Industrial Engineering and Management, singly or in small groups. Melons do not ripen at the same time. Each Ben Gurion University of the Negev, Beer Sheva 84105, Israel. G. E. Miles is with the Department of Agricultural Engineering, Purdue fruit must be detected separately and evaluated for ripeness prior to the harvesting. Fruit picking with a sensor-based robot that emulates University, West Lafayette, IN 47907 USA. IEEE Log Number 9402292. the human picker seems a viable potential solution for automation of 0018-9472/94$04.00 0 1994 IEEE

I

1260

IEEE TRANSACTIONS ON SYSTEMS,MAN. AND CYBERNETICS, VOL. 24, NO. 8, AUGUST 1994

evaluated were average cycle times and number of fruit successfully picked.

B. Database

I ll.

Fig. 1. Mechanical drawing of the robotic melon harvester.

melon harvesting which is currently a laborious and expensive task [191, [231.

n.

OVERVIEW OF ROBOTIC MELON HARVESTER OPERATION

The robotic field crop machine consists of a manipulator mounted on a rectangular mobile chassis which is drawn by a tractor [3]. Fruit are picked while the vehicle advances along the row. The chassis is divided into two areas (Fig. 1): the front area is a shaded, unobstructed space which provides a clear field of view for detection of oncoming fruit. The arm’s workspace is located right behind it. The front area consists of a machine vision system (far-vision) which identifies two dimensional locations of the fruit lying ahead of the arm. It provides input for planning the robot’s motions and highlights areas of interest for a second vision system (near-vision) which is located on the robot manipulator and guides the arm toward the fruit. The robot arm approaches each fruit and detects the ripes ones using a volatile gas ripeness sensor [4]. Ripe fruit are picked using a custom designed gripper [30] and placed on discharge conveyors. Real-time image processing algorithms have been developed and implemented to locate melons based on reflectance, shape and size [9]. During the picking, the far-vision operates cooperatively with an oscillating air blast which blows away the foliage and exposes melons which are hidden by the canopy. An intelligent control system has been structured based on the blackboard model for real-time operation of all sensors and actuators [13].

III. METHOD

A. Design Methodology An initial conceptual design was derived by comparing alternative designs through graphic simulations [ 141 which provided insight into the systems problematics and outlined feasible design alternatives. However, due to the many feasible horticultural practices and variability of crop conditions not all possibilities could be predicted and compared using graphic simulation. Therefore, numerical simulation tools were developed and applied by developing computer programs that modelled the system dynamics in Fortran. Alternatives were compared statistically considering all crop parameters such as spatial geometry of the crop and fruit distribution. Analysis of the many coupled design parameters was achieved by running an extensive set of simulations where in each simulation test only the values of one parameter was varied, while all the other parameters were held constant. The effect of varying the tested parameter was analyzed for many combinations of all the other parameters. Thus, only the influence of the dependent variable was analyzed. Each design parameter was varied in a similar manner. From the derived relationship patterns, the best set of parameters was selected. The performance criteria

-

7-

Since distribution of the fruits along the row effects the design and performance of an agricultural robot [121 actual field coordinates of fruit in typical rows must be measured and serve as input to the simulation model. In addition, machine performance should be evaluated for different common horticultural practices to provide guidelines to adapt the specific crop conditions that yield the best results. A methodology how to analyze the data and how to generate feasible datasets was developed [14]. Since this is crop dependent it must be modified according to the horticultural practices of the specific crop evaluated. For the robotic melon harvester, field locations of cantaloupes, cultivar ‘Superstar’, were collected on 859 melons in 29 rows [14]. These data served as input to all simulations performed. To generate new datasets that fit common planting distances, distribution curves were fitted to statistics derived from the measured locations, e.g., average distance between fruit, average distance between plants, number of fruit clustered near the plant center. The Chi-square test was used to test the goodness of fit [29]. Similar datasets were generated by changing plant distances between 25 and 125 cm (which are common planting distances for melons [5]) and fitting the data to the best fit data curve (the one with the lowest Chi-square value). The number of fruit per plant depends on the planting distance and was computed according to the equations derived by Bhella [SI. For each ‘simulated’ planting distance (denoted as Dz), three different random seed numbers were used to generate the data. For each random seed number, a total of lo00 fruit coordinates were generated using the inverse transformation method [24]. Additional data on melon locations was generated assuming uniform distribution across the row to test the influence of the distribution pattern of the melons on the performance criteria. C. Design Parameters Type ofRobot Comparison of two types of robots through graphic simulation indicated that the Cartesian robot was preferable [14]. To decide on the best performing robot in a more systematic way, previous algorithms that minimize the total operational time by determining the optimum sequence ([l], [6], [lo], [ l l ] , [12]) were expanded to compare different robots and find the robot that achieves the global minimum time: For each type of robot all possible paths from all possible placements of the robot base were searched. The costs of the minimum-time path for each robot were compared. The motion sequence was obtained by solving the traveling salesman problem [8]. By comparing the minimum travel times obtained for different robots, the most time-efficient robot was selected. For the melon harvesting task, comparisons were performed for a cylindrical and a Cartesian planar robot. All actuator speeds were assumed to be 1 c d s e c . The minimum and maximum link lengths were 30 and 120 cm. An extensive search through all possible paths and base locations of the robot was performed to determine the minimum-time sequence in which to pick the fruit. Workspace Design The robot’s workspace depends on the horticultural practices and should enable access to any location across the growing bed and to conveyors for discharging the picked fruit. Hence, the conveyors location must also be determined. Two different conveyor configurations were evaluated: the first consisted of two conveyors, each one located on the side of the bed, along the row, in the direction of the tractor’s motion (X). The second conveyor configuration evaluated was located across the row at the rear end of the robot’s workspace (Y).

lEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS, VOL. 24, NO. 8, AUGUST 1994

To enable simultaneous picking of the fruit while the tractor is advancing at a constant speed along the row (X), the manipulator must have a degree of freedom in the vehicle’s direction of motion. This allows a greater selection in the point of interception with the fruit and compensates for nonuniform fruit distribution. The following values of the longitudinal degree of freedom (X) were simulated: 0.2, 0.4, 0.6, 0.8, 1, 1.2, 1.4, 1.6, 1.8, 2.0, 4.0, 6.0,8.0 m. Dynamic Characteristics The performance of the robotic field crops machine depends on its dynamic characteristics such as actuators speeds and forward speed of the tractor. Increasing the actuator’s velocity will increase productivity, however the exact influence was determined by measuring average cycle time for different speeds of both the transverse (V,) and longitudinal (V,) actuators and different tractor speeds (Vt). To simplify the simulation process, an additional variable ( V f )which combines the speeds in the tractor’s direction of motion (X) was defined:

1261

direction of motion 4

(b) direction of motion 4

v, = v, + vt. The following speeds were simulated: V,, V,: 11.8,25,50, 100, 200, 400, 800, 1600 c d s e c ; V t : 11, 22, 33, 66 c d s e c (0.4, 0.8, 1.0, 2.0 km/hr correspondingly). Based on field experiments with the prototype gripper the picking time ( P t ) was estimated as 1.5 seconds. However, to evaluate its effect on the overall average cycle time, picking times of 0.5, 1.0, 1.5. and 2.0 were simulated. Number of Arms The robots could be more fully utilized by reducing the number of arms, although the total time would increase. Altematively, the total time could be decreased by increasing the number of robots [18]. The tradeoff between the number of systems and their performance was determined systematically. Average cycle times were determined for up to six arms working in parallel across the row, assuming non overlapping workspaces and therefore no need to deal with collision avoidance. Average cycle times were determined by dividing the total time by the total number of fruit harvested by all arms. This was derived for different speeds of the transverse (V,) actuator and different forward ( V f )speeds for both measured field data and uniformly distributed data. Multiple Arm Conjiguration Various operational modes of multiple arms have been described in industrial applications [21]. The configurations considered for the agricultural domain were: Single: one arm harvesting all the fruit growing across the bed (Fig. 2(a)). Parallel: two identical arms doing the same task, each harvesting the fruit on half a bed and conveying the fruit to a nearby conveyor (Fig. 20)). Sequential: two arms, the first arm harvests the fruit and transfers it to the nearby second arm that transfers the picked fruit to the main conveyor (Fig. 2(c)). Tandem: two identical arms acting in a tandem mode. One is following the other. Each arm is responsible for picking and conveying the fruit (Fig. 2(d)). The average harvest time for each configuration is calculated from the total time divided by the number of fruit successfully picked. IV. DESIGN OF A ROBOTIC MELON HARVESTER

A. Melon Location Analysis The spatial distribution of the fruit was statistically analyzed using the International Mathematical and Statistical Library (IMSL Inc.). The average distance between adjacent fruit along the row (X) was 34 cm with a standard deviation of 10 cm. The average

Fig. 2. Schematic diagram of the different configurations:(a) 1 arm; (b) 2 parallel arms; (c) 2 sequential arms; (d) 2 tandem arms.

distance between fruit across the row (Y) was 23.6 cm with a standard deviation of 14.2 cm. To generate additional datasets for the simulations, distribution curves were fitted to the histograms of the distance between clusters of fruit and distance fom the cluster centroid. Clusters of fruit were determined by a hierarchical cluster analysis using Ward’s method [2]. Detailed description of the statistical analysis can be found in [14].

B. Type of Robot The Cartesian robot was faster for 266 of the 276 cases 1141. Moreover, for 34 cases out of the 276 tested, not all fruit were reachable from all possible locations of the robot base in the fruit cluster area for a cylindrical robot. This means, that in order to pick all fruit, the cylindrical robot sometimes will need to move to different base locations. The Cartesian robot can reach all fruit from the central base location without moving. C . Workspace Design, Task limes and Tractor Speeds

Locating the conveyors along the row decreases the average cycle time (Fig. 3). The cultivated row width was 120 cm and hence the workspace of the horizontal axis (Y) must be be wider and was recommended to be 150 cm to enable the robot to reach the discharge conveyors which are positioned along the chassis on both sides. The vertical clearance of the gripper arm was recommended to be 90 cm which is enough to pick the fruit and load it to a conveyor. The optimal work range of the X actuator (the frame size) was determined through simulations for different picking times and frame sizes ([13], [14]) which indicated the following: 1) As the frame size increases until 0.8 m, the percentage of fruit picked increases. For frame sizes between 0.8 and 2 m the percentage of fruit picked does not change, except for a slight increase for picking times of 0.5 seconds and 2 seconds at a frame size of 1 m.

1262

IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS, VOL. 24, NO. 8, AUGUST 1994 1

:j

......................................................................................... .........................................................................

'5

ae

-............-. ...........-.

-..-_.-...... "......-...... ....................................................................

4-

44 .E

f

,8,

8

4

5

6

NunkrdPrms

Fig. 5. Utilization of multiple arms using field data. Fig. 3. Comparing percent time difference for different conveyor configurations for a Cartesian robot.

ir"i

E

t

Number d atnu

"rdarmr

Fig. 6. Effects of number of manipulators and actuator speeds on harvest cycle time for a forward speed of 100 cdsec.

Fig. 4. Utilization of multiple arms using uniformly distributed melons. 1

2 ) For frame sizes larger than 2 m there is no significant increase

in the fruit percentage successfully picked for the different picking times and tractor speeds evaluated. 3) For a tractor speed of 33 c d s e c the percentage of fruit picked for frame sizes up to 2 m remains constant and beyond that, decreases. 4) For slow tractor speeds of 11 and 2 2 c d s e c the percentage of fruit picked increases for frame sizes up to 100 cm, then stays constant for frame sizes from 1 to 2 m and then decreases for frame sizes greater than 2 m.

1

-H' = 1

1

F: EI-

D. Number of Arms and Actuator Speeds For the uniformly distributed data, each arm harvests the same number of fruit and is utilized the same percentage of time regardless of position across the melon bed (Fig. 4). The utilization of each arm decreases as the number of arms increase. For field data, the center arms are heavily utilized because fruit are clustered in the middle (Fig. 5). Therefore, adding arms does not significantly increase machine capacity when each robot is restricted to harvesting the same width of the row. The effect of number of manipulators on the average time to harvest a melon for forward speeds of 100 c d s e c and 25 c d s e c is illustrated in Figs. 6 and 7. Increasing the number of manipulators decreases the

Number d arms

Fig. 7. Effects of number of manipulators and actuator speeds on harvest cycle time for a forward speed of 25 cdsec. time per fruit; however, the rate of decline decreases as the number of arms increase and is less than 1% for more than three arms. The influence of the transverse actuator speeds decreases as the number of arms increase. Comparing Fig. 6 to 7 reveals that increasing forward speed decreases the time per fruit more significantly than increasing the transverse actuator speed.

U-M-I DUE TO LACK OF CONTRAST,GRAPHS DID NOT REPRODUCE WELL. GRAPHS FOLLOW S A M E SEQUENCE AS LEGEND

I

-

-

-

1263

IEEE TRANSACTIONSON SYSTEMS, MAN, AND CYBERNETICS, VOL. 24. NO. 8, AUGUST 1994

:1 za

i

F

T

2

1

3

4

5

do

6

20

NUIllbCKofPnrc

Fig. 8. Effects of number of manipulators and actuator speeds on harvest cycle time for a transverse speed of 100 cdsec.

40

I

€4 80 100 120 140 160 A c w speed almg the row (cmhsc)

m

Fig. 10. Effects of configuration and longitudinal actuator speeds on harvest cycle time for a transverse speed of 11.8 cdsec using field data. 30

..*..

]I

1% -

2-

Y

P-

180

‘2

i

=I

k

l5-

p E lo.

E

F

1

2

3 4 h*dm

5

6

Fig. 9. Effects of number of manipulators and actuator speeds on harvest cycle time for a transverse speed of 25 cdsec.

Fig. 11. Effects of configuration and longitudinal actuator speeds on harvest cycle time for a transverse speed of 100.0 cdsec using field data.

The effect of number of manipulators on the average harvest time for transverse actuator speeds (Y axis) of 100 c d s e c and 25 c d s e c is illustrated in Figs. 8 and 9. Increasing forward speeds significantly decreases the average cycle time. When the number of arms is greater than 3, changing the transverse actuator speed from 25 to 100 cm/sec speed has little influence on cycle times.

E. Multiple Arm Configuration and Actuator Speeds The tandem configuration was the fastest followed by the parallel, sequential mode and a single arm respectively for different forward speeds evaluated for a transverse actuator speed (V,) of 11.8 c d s e c (Fig. 10) and 100 c d s e c (Fig. 11). For all cases, increasing the X actuator speed decreases the cycle time. For V, greater than 200 c d s e c , the cycle time declines gradually. As V, increases the difference between the configurations becomes smaller. The tandem configuration was the fastest for different transverse actuators speeds (Y) evaluated for forward speeds of 12.5 and 62.5 c d s e c (Figs. 12 and 13). For all cases, increasing the Y actuator speed reduces the cycle time. As V f increases the difference between the configurations becomes smaller. Transverse actuator speeds exceeding 50 c d s e c have little effect on cycle times. Comparing Figs. 12 and 13 indicates that increasing the longitudinal actuator speed,

$

40

60

80

loo

120

Horiununlac-spted(cm/sec)

140

160

180

m

Fig. 12. Effects of configuration and forward speeds on cycle time for a longitudinal speed of 12.5 cdsec using field data. decreases the required cycle time more significantly than increasing the transverse actuator speed. Additional simulations [141 indicated the following: 1) Less than 65 percent of the fruit is picked when tractor speed exceeds 22 c d s e c (0.8 km/hr) and only one arm with a 1.5 second picking time is used.

U-M-I DUE TO LACK OF CONTRAST,GRAPHS DID NOT REPRODUCE WELL. GRApas FOLLOW SAME SEQUENCE AS LEGEND 1

I

20

IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS,VOL. 24, NO. 8, AUGUST 1994

1264

34

'0

1

2o

w, m

4o

iw

i2o

iu)

iw

i8o

I

m

H"JrdusDrspwd(-)

Distance betwean pllants (Cm)

Fig. 13. Effects of configuration and fonvard speeds on cycle time for a longitudinal speed of 62.5 c d s e c using field data.

Fig. 14. Effect of planting distance on cycle times for the tandem configuration and for 2 arms acting on conventional and experimental planting distances.

2) For a slower tractor speed of 11 c d s e c (0.4 k m h ) , and picking time of up to 2 seconds, more than 80 percent of the fruit can be successfully picked with one arm. Two arms are necessary if the tractor speed is to be increased. 3) Two tandem arms pick 90 percent of the fruit with a task time of 1.5 seconds and tractor speed of 22 c d s e c (0.8 kmhr). If task time is increased to 2 seconds only 80 percent of the fruit are picked for the same conditions.

an extensive analysis which included a wide range of values of all the design parameters, characteristic performance patterns were generated. Based on these performance curves, preferable parameters were recommended. The applicability of the developed systems engineering methods was demonstrated by using it to determine design parameters of a robotic melon harvester. Performance of a robotic melon harvester was simulated to quantify effects of type of robot, number of arms, multiple configuration, and actuator speeds. An algorithm developed for comparing different robots indicated that the Cartesian robot performed faster and was capable of reaching all fruit. Additional advantages of the Cartesian robot such as easier control, better dynamic performance and the fact that it usually has higher resolution with equal and constant spatial resolution makes it preferable. Numerical simulation on measured field data indicated that increasing the number of arms beyond two does not reduce the cycle time significantly. This is because the fruit are clustered around the center of the plant and arms harvesting the sides of the row were underutilized. Therefore, activating two arms in tandem was the fastest configuration. Even for uniformly distributed melons, adding more than three arms does not reduce time per fruit significantly. Based on the simulations performed, the recommended actuator speeds for the robotic melon harvester should be 100 c d s e c . Higher speeds will only slightly reduce the cycle time. Two arms are necessary to increase the tractor speed. Two tandem arms can succeed in picking 90 percent of the fruit with a picking time of 1.5 seconds, and tractor speed of 22 c d s e c (0.8 km/hr). The workspace should be: Y: 1.5 m; 2: 0.9 m and X: between 1 and 2 m. Larger workspace in the X direction will result in decrease of percentage of fruit successfully picked, since the arm will have to travel longer distances. Smaller workspace in the X direction cause many fruit to be missed. Based on these simulations a robotic melon harvester has been constructed (Fig. 15). Since, each agricultural robot must be custom-designed and adapted to the specific crop, the design methodology presented in this work is important for evaluating new applications before expensive prototypes are built and field-tested. The methodology presented can be applied to determine performance for proposed applications and thus substantially reduce the research and development costs associated with new agricultural robots. The work presented in this paper is an essential prerequisite for the successful development of a mobile, field robotic harvester which will undoubtedly be a breakthrough for mechanization of highly perishable agricultural products.

F. Planting Distances All the aforementioned analyses were evaluated for planting distances between 25 cm and 125 cm [13]. Simulation results indicated that cycle time decreases as the distance between plants increase. The best performance was achieved for tandem arms since most of the fruit are concentrated along the center of the row. Hence, it was assumed that performance could be further improved by operating two parallel arms if the number of melons on each half row were were uniform. To achieve an equal number of melons for each arm an experimental growth pattern which consisted of alternating plants, each plant centralized in the middle of the half bed (instead of sequential plants along the middle of the whole bed) was proposed, simulated and field tested [ 141. Simulation results for two arms picking simultaneously, each on half a row, on this proposed horticultural design indicated that the average cycle time is 1.31 seconds per fruit which is a 65.8% reduction from the tandem configuration on the conventional cultural practice, for a picking time of 1.5 seconds (Fig. 14). Since 2 fruit are picked concurrently, the total time that accumulates is the maximum time between the two systems and the time per fruit is the total time divided by the number of fruit. This alternative spacing favors two parallel arms since now each arm has less transverse distances to travel.

v.

SUMMARY AND CONCLUSION

This work presents a general approach for designing agricultural robots. A systems engineering method has been developed and demonstrated to determine the many closely related design parameters of a robotic system performing in unstructured environments. Graphic simulation initiated the design process and permitted assessment of alternative concepts and designs in a timely manner. To optimize performance based on statistical and sensitivity analysis, numerical simulation tools were developed and applied. To simplify the simulation process, the effect of each design parameter was separately evaluated while holding constant all other values. Through

IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS, VOL. 24, NO. 8, AUGUST 1994

1265

Automation and Applications, John Wiley and Sons, Inc., New York,

Fig. 15. Photograph of the robotic melon harvester constructed basedon the systems engineering analysis. ACKNOWLEDGMENT We would like to acknowledge Deneb Robotics Inc. and IMSL Inc. for the analysis derived from their software systems.

REFERENCES L. L Abdel-Malek and Z. Li, “Application of inverse kinematics in the optimum sequencing of robot tasks,” International Journal of Production Research, vol. 28, no. 1, pp. 75-90, 1990. M. R. Andersberg, “Cluster analysis for applications. Academic Press,” New York, 1973. M. Benady, Y. Edan, A. Hetzroni and G. E. Miles, “Design of a field crops robotic machine,” ASAE Paper No. 91-7028, ASAE, St. Joseph, MI 49085, 1991. M. Benady, J. E. Simon, D. J. Charles and G. E. Miles, “Determining melon ripeness by analyzing headspace gas emissions,” ASAE Paper No. 9 2 6 0 5 5 , ASAE, St. Joseph, MI 49085, 1992. H.S. Bhella, “Response of muskmelon to within-row plant spacing,” Proceedings of the Indiana Academy of Science, vol. 94, pp. 99-104, 1984. J. Borenstein and Y. Koren, “Task-level plan generation for mobile robots,” IEEE Transactions on Systems, Man and Cybem., vol. 20, no. 4, pp. 938-943, 1990. C. Chen and R.G. Holmes, “Simulation and optimal design of a robot arm for greenhouse/nursery,” ASAE Paper No. 90-1539, ASAE, St. Joseph, MI 49085, 1990. N. Christofides, “The Traveling Salesman Problem,” in: Graph Theory-An Algorithmic Approach, pp. 236281. Editor: w. Rheinbolt. Academic Press, New York, 1975. Y. Dobrusin, Y. Edan, J. Grinspun and U. M. Peiper, “Real-time image processing,” ASAE Paper No. 92-3.515, ASAE, St. Joseph, MI 49085, 1992. Z. Drezner and S. Y. Nof, “On optimizing bin picking and insertion plans for assembly robots,” IIE Transactions, vol. 16, no. 3, pp. 262-270, 1984. S. Dubowsky and T. D. Blubaugh, “Planning time-optimal robotic manipulator motions and work places for point-to-point tasks,” IEEE Transactions on Robotics and Automation, vol. 5, no. 3, pp. 377-381, 1989. Y. Edan, T. Flash, U. M. Peiper, Y. Sarig and I. Shmulevich, “Near minimum-time task planning for fruit-picking robots,” IEEE Transactions on Robotics and Automation, vol. 7, no. 1, pp. 48-56, 1991. Y. Edan, “Control and Design of an Intelligent Agricultural Robot,” Ph.D. Thesis, Purdue University, W. Lafayette, IN 47907, 1990. Y. Edan and G. E. Miles, “Design of a robotic melon harvester,” Transactions of ASAE, vol. 36, no. 2, pp. 593-603. 1993. C. A. Klein and A. A. Maciejewski, “Simulators, Graphic,” in, R. C. Dorf and S. Y. Nof, eds., International Encyclopedia of Robotics,

1991. [16] Y. Koren, Robotics for Engineers, McGraw Hill Book Company, New York, 1985. [17] L. J. Kutz and G. E. Miles, “CAD simulation speeds robotic workcell development,” Agriculture Engineering, May/June, pp. 20-22, 1986. [18] C. S. G. Lee, “Robot Mechanisms: Algorithms and Architectures,” Personal Communication, Electrical Engineering, Purdue University, West Lafayette, IN 47907., 1990. [19] D. H. Lenker, “Factors limiting the harvest mechanization of some major vegetable crops in the U.S.A.,” in Fruit, Nut, Vegetable Harvesting Mechanization Symposium, pp. 29-38, 1984. [20] G. E. Miles and Y. Tsai, “Combine systems engineering by simulation,” Transactions of the ASAE, vol. 30, no. 5, pp. 1277-1281, 1987. [21] S. Y. Nof, Handbook of Industrial Robotics, John Wiley and Sons, New York, 1985. [22] S. Y. Nof, “Industrial Robotics,” in, G. Salvendy, ed., Handbook of IE, John Wiley and Sons, Inc., New York, 1991. [23] M. O’Brien and M. Zahara, “Mechanical harvest of melons,” Transactions of the ASAE, vol. 14, no. 5, pp. 883-885, 1981. [24] A. B. Pritsker, Introduction to Simulation and Slam I!, John Wiley and Sons, New York, 1986. [25] E. I. Rivin, Mechanical Design of Robots, McGraw Hill Book Company, New York, 1988. [26] Y. Sarig, “Robotics of fruit harvesting: a state-of-the-art review,” Journal of Agricultural Engineering Research, 54, pp. 265-280, 1993. [27] K. C. Ting, Y. Yang and W. Fang, “Stochastic modeling of robotic workcell for seedling plug transplanting,” ASAE Paper No. 90-1 539, ASAE, St. Joseph, MI 49085, 1990. [28] A. A. Tseng, “Assessment of robotic simulation software,” Proceedings of the I988 ASME Computers in Engineering Conference, vol. 2, pp. 217-226, 1988. [29] R. E. Walpole and R. H. Meyers, “Probability and statistics for engineers and scientists,” McMillan Publishing Co., New York, 1985. 1301 I. Wolf, I., J. Bar-Or, Y. Edan and U. M. Peiper, “Developing grippers for a melon harvesting robot,” ASAE Paper No. 90-7504, ASAE, St. Joseph, MI 49085, 1990.

Acceleration Based Learning Control of Robotic Manipulators Using a Multi-Layered Neural Network K. H. Kyung, B. H.‘Lee and M. S. KO

Abstract-This paper presents a nonlinear compensation method based on neural networks for trajectory control of robotic manipulators. A multi-layered perceptron neural network (MLP) is used to predict the required actuator torques of a robot to follow a desired trajectory, and these predicted torques are applied to the robot as feedforward compensations in parallel to a linear feedback controller. An acceleration based learning scheme is proposed to adjust the connection weights in the neural network to form an approximated dynamic model of the robot. Simulation results show that the proposed learning scheme improves the speed of error convergence of the system and reduces the convergent error with the efficient adaptation to the changing system dynamics. The validity of the proposed learning scheme is verified through experiments.

Manuscript received October 3, 1992; revised September 10, 1993. K. H. Kyung and B. H. Lee are with the with the Department of Control and Instrumentation Engineering, Seoul National University, san 56- 1, Sinrim 2 Dong, Kwanak Gu, Seoul, 151-742, Korea. M. S. KO is professor emeritus of the Department of Control and Instrumentation Engineering, Seoul National University. IEEE Log Number 9402291.

0018-9472/94$04.000 1994 IEEE