DISTRIBUTED MULTI-ROBOT ASSEMBLY/PACKAGING ALGORITHMS

Transcription

1 Intelligent Automation and Soft Computing, Vol. 10, No. 4, pp , 2004 Copyright 2004, TSI Press Printed in the USA. All rights reserved DISTRIBUTED MULTI-ROBOT ASSEMBLY/PACKAGING ALGORITHMS Y. EDAN, S. BERMAN 1, E. BOTEACH, E. MENDELSON Dept. of Industrial Engineering and Management Ben-Gurion University of the Negev Be`er Sheva, Israel ABSTRACT Five different algorithms for cooperation between assembly/packaging robots in a distributed system were developed and evaluated in simulation. Independent, autonomous operation of each robot was assumed. The goal was to maximize the number of assemblies while minimizing the standard deviation of the work between the different robots. Two of the algorithms, a fuzzy algorithm and a simple take-what-you-can algorithm were implemented on an experimental system. Experimental results indicated better performance for the simple algorithm in terms of the number of assemblies. However, the fuzzy algorithm reduces the standard deviation of work between the robots and therefore equalizes the robots utilization. Equal utilization is a major factor in multirobots systems, and contributes in decreasing mean time between failure (MTBF). 1. INTRODUCTION Key Words: Multi-robots, fuzzy logic, behavior-based, assembly Assembly systems are characterized by a stochastic nature [17] which includes uncertainty in parts arrival time, dimensions of manufactured parts, and parts placed at random positions and orientations. Therefore, to enable industrial robotization of assembly tasks, capabilities to operate in an uncertain world must be developed. This requires developments of robots that synthesize their own manipulation sequences and invoke sensors and fixtures only when needed, robots with abilities to deal with unexpected events. Robotic workstations employing two or more robots can be advantageous when resource and task sharing lead to reduced costs and improved production rate [14]. Cooperation between robots can save time, resources, and space while enabling complex tasks [5] and is considered an essential attribute of intelligent manufacturing systems [16]. However, algorithms must be developed for efficient and effective cooperation and coordination. Extensive research has been conducted on synchronizing multi-robots in assembly, using optimal sequencing procedures, petri nets, developing hierarchical activities controllers and applying time-buffer methods to balance multiple arms [2, 14, 21]. However, the application of these approaches in flexible manufacturing systems has been very limited considering the dynamic nature of the assembly work cell and its inherent uncertainties. Behavior-based control enables a robot to perform in an unknown dynamic changing environment [2, 7]. This approach provides fast and robust action in the absence of an explicit plan. The main advantage is that no modeling is required since the basic behaviors react to stimuli from the environment. This enables implementation in unstructured environments. Such systems have been widely applied in mobile robots [10, 18, 19]; and have also been applied to robot manipulators [9] though to a much lesser extent [20]. Another mathematical tool for handling real world uncertainty is fuzzy logic [11]. It provides a method for translating qualitative and imprecise information into quantitative terms. Several researchers have applied fuzzy logic to behavior-based systems for multi-robot control. Most applications are in mobile robots in which fuzzy logic is used for behavior amalgamation [4, 10]. One of the key features in the design of a multi-robot system is its control methodology, i.e., how conflicts are resolved and how the joint capabilities of the multi-robot apparatus are utilized [14]. The most 1 Currently with Weizmann Institute of Science, Israel 349

2 350 Intelligent Automation and Soft Computing fundamental decision to be made is whether the system is centralized or distributed. (i.e., decentralized). Centralized systems are controlled by a single control entity. This may be infeasible for complex largescale systems. Such systems may be too complex for traditional control algorithms, spatially scattered, or have a multitude of interactions between sub-systems [9]. Decentralized operation is expected to contribute to system flexibility [5, 11], robustness and modularity [1, 3] however, there is a need to coordinate between the robots. The objective of this research was to develop and demonstrate algorithms for a distributed multi-robot assembly/packaging system. 2. METHODOLOGY The algorithms were applied to an assembly/packaging task with multi-robots, each of which has the same control algorithm. Independent, autonomous operation of each robot was assumed with nonoverlapping workspaces. The goal of each robot was to assemble/package a part that is a combination of two types of sub-parts. There was no constraint on the assembling/packaging sequence. The following parameters were varied in a simulation which compared five different algorithms: the number of robots; location of robots; robot cycle times, and assembly configuration (i.e., the number of sub-parts from each type for the assembly/packaging of a full part). The performance criteria were defined as the number of assemblies and distribution of workload between the robots. The goal was to maximize the number of assemblies (denoted as NUM_ASBY) while minimizing the standard deviation of the work between the different robots (denoted as STD_BR). The algorithm efficiency measure was defined as the weighed combination of these two parameters and was normalized to facilitate comparison between different simulated conditions [eq. 1]. Algorithm efficiency measure=w1 STD_BR+w2 NUM_ASBY (1) where: w1, w2 are the weights of each parameter An experiment was performed to validate the simulations and included implementation of two of the simulated algorithms the simplest algorithm and the most complicated one. 3. SIMULATION 3.1 Description A simulation model was developed in Visual C++ and consisted of the following elements: a conveyor, two types of sub-parts entering the conveyor, and several robots on each side of the conveyor. The following parameters could be varied [6]: the number of robots, robots location along the conveyor, robot cycle times, conveyor cycle times and the number and types of sub-parts constituting a package. The following algorithms were developed, implemented and evaluated [6]: Simple: implements the take-what-you-can principle. If a sub-part is needed to complete the assembly and available on the conveyor - the robot takes it. This algorithm is simple to implement and does not require any data collection or sharing. It is used as the baseline for comparison to the other algorithms. Average: This algorithm ensures equal utilization of all robots: each robot assembles new parts only if its number of assemblies is less than the average number of assemblies. Each robot must know the other robots average number of assemblies. Occupancy: This algorithm tries to reduce the need to wait for the slower robots when robots with different cycle times are employed. A utilized step is defined as a step in which the robot performed an operation; a non-utilized step is when the robot did not perform an operation. This algorithm tries to equalize the relation between non-utilized and utilized steps for all robots. Behavior-based: This algorithm strives to equalize robot utilization without the need of explicit knowledge about other robots. The reactive nature of the algorithm is suited for the stochastic nature of the parts arrival rate. Two behaviors were defined: satisfaction (high productivity), frustration (low productivity). The output is a decision to take or leave the current sub-part. The robot takes a part whenever it cannot perform an assembly aiming to maximize productivity. Fuzzy: This algorithm tries to maintain equal utilization under uncertainty and perturbations in the part arrival rate. The fuzzy output, defined as a decision to take or leave a sub-part takes into account the

3 Distributed Multi-Robot Assembly/Packaging Algorithms 351 following fuzzy variables: Assembly ratio - the ratio of robot assemblies relative to the average number of assemblies; Utilization ration - the ratio of robot utilization relative to the average utilization; and Location - the robot s location. These parameters are weighted using eight fuzzy rules (Figure 1). The robot reacts according to its condition, and decides whether to pick the incoming part or to rest in current control cycle. This algorithm considers many parameters and therefore, is considered a complicated algorithm for implementation. Assembly ratio Assembly ratio Low close_to_ average Take part Take part High Medium Assembly ratio high Take part Low IF Utilization ratio low Take part High Utilization ratio IS close_to_ THEN Take part IS Medium average Utilization ratio High Take part Low Location Close Take part Medium Location Far Take part High Figure 1. Fuzzy rule base Analysis of the algorithms were conducted for different number of robots, different robot locations along the conveyor, different robot cycle times and different configurations of the package (number of subparts from each type). In each simulation only one parameter was varied. The following were considered as the base line values: 6 robots, conveyor and robot cycle times of 1000 ms, package configured of 4 subparts of type a and 2 sub-parts of type b. Each simulation consisted of 1000 sub-parts that entered the conveyor randomly. 3.2 Results Results (Table I) indicated that for all cases the simple algorithm yielded the maximum number of assemblies (NUM_ASBY) but with the maximum differences between the robots (maximum STD_BR). The average algorithm resulted in equal performance of all robots (i.e., STD_BR was zero) as expected, but with an overall low NUM_ASBY. The second best performing algorithm regarding NUM_ASBY and STD_BR were the fuzzy and occupancy algorithm respectively. In most cases the fuzzy algorithm achieves lower STD_BR than the behavior-based algorithm. The robots locations hardly influences the NUM_ASBY in the simple, average, occupancy and fuzzy algorithms and had small effect on the outcome also in the behavior algorithm. Differences in cycle times have significant influence on both NUM_ASBY and STD_BR and caused differences of approximately 40% in NUM_ASBY for the average algorithm as compared to only 20% differences in all other algorithms. In all algorithms when the assembly/package is composed of equal distribution of sub-parts the highest NUM_ASBY is achieved. Table II describes the efficiency measure for different weights of the NUM_ASBY and the STD_BR. For all algorithms, except for the simple algorithm, the efficiency measure decreases as the relative weight of the NUM_ASBY increases. For the simple algorithm, as expected, the trend is vice versa (since the algorithm gives priority to maximum NUM_ASBY). In all cases (except for case in which the relation between STD_BR and NUM_ASBY is 35-65) the fuzzy algorithm resulted in the best efficiency. The

4 352 Intelligent Automation and Soft Computing Table I. Simulation results* *Total number of assemblies (standard deviation between the different robots) (a) Number of robots package consists of 4 parts of type a and 2 parts of type b; robot and conveyor cycle times = 1000ms Number of Simple Average Occupancy Behavior Fuzzy robots 3 99 (5.2) 84 (0) 78 (0) 71 (1.15) 83 (0.58) (8.62) 92 (0) 88 (0) 89 (2.99) 95 (0.5) (15.63) 108 (0) 101 (0.41) 101 (2.40) 111 (1.22) (11.56) 120 (0) 110 (0) 62 (0.42) (2.62) (11.56) 120 (0) 100 (0) 60 (0) 117 (2.7) (b) Robots location along the conveyor 6 robots; package consists of 4 parts of type a and 2 parts of type b robot and conveyor cycle times = 1000ms Location* Simple Average Occupancy Behavior Fuzzy 1 (one side) 120 (15.54) 107 (0.41) 101 (0.41) 108 (0.0) 111 (2.26) 2 (4:2) 120 (15.54) 108 (0) 102 (0) 101 (2.4) 109 (1.61) 3 (3:3) eq. 121 (15.63) 108 (0) 101 (0.41) 101 (2.4) 111 (1.22) 4 (3:3) al. 120 (15.54) 107 (0.41) 101 (0.41) 101 (1.94) 111 (1.38) 5 (3:3) al.,eq.,al., eq 120 (15.23) 108 (0.0) 101 (0.41) 92 (1.63) 110 (2.16) Location* (X:Y indicate number of robots on each side of conveyor, X left side, Y right side; Eq. Indicates one opposite the other, Al. Indicates alternating locations: one robot on one side with no robot opposite it, second robot on other side with no robot opposite it, etc.) robot conveyor motion direction

5 Distributed Multi-Robot Assembly/Packaging Algorithms 353 (c) Robot cycle times (X:Y:Z;R:S:T indicate robot cycle timex1000ms; conveyor 1000ms) 6 robots, 3 on each side; package consists of 4 parts of type a and 2 parts of type b Cycle times* Simple Average Occupancy Behavior Fuzzy (1:2:3; 2:3:1) 104 (10.76) 60 (0) 83 (7.28) 86 (3.83) 80 (3.14) (3:2:1; 3:2:1) 104 (6.25) 66 (0) 80 (6.95) 89 (7.29) 86 (3.14) (2:2:2; 2:2:2) 121 (15.63) 108 (0) 101 (0.41) 101 (2.4) 111 (1.22) (1:2:3; 1:2:3) 105 (14.92) 52 (0.52) 65 (5.67) 72 (4.05) 63 (2.95) (3:1.5:1; 110 (10.44) 66 (0.0) 82 (5.13) 81 (2.43) 81 (3.99) 1.5:1:2.5) Cycle times* (X:Y:Z ; R:S:T indicate relationship between cycle times of robots on each side of conveyor Where X,Y,Z are located on left side and R,S,T are located on right side of conveyor, e.g.1:2:3; 2:3:1 indicates robots with following cycle times 1000:2000:3000; 2000:3000:1000) Z Y X T S R (d) Assembly configuration (number and type of sub-parts: Xa, Yb: X parts of) 6 robots, 3 on each side robot and conveyor cycle times = 1000ms Assembly configuration Simple Average Occupancy Behavior Fuzzy 1 (2a, 2b) 235 (15.21) 174 (0) 180 (0) 180 (2.28) 189 (0.84) 2 (1a, 3b) 163 (23.55) 136 (0.52) 143 (0.41) 151 (6.74) 148 (2.88) 3 (4a, 2b) 121 (15.63) 108 (0) 101 (0.41) 101 (2.4) 111 (1.22) 4 (4a, 0b) 121 (21.09) 108 (0) 106 (0.52) 111 (4.93) 111 (1.76) 5 (7a, 3b) 69 (8.14) 57 (0.55) 54 (0.0) 48 (0.63) 63 (1.22) Table II. Summary of algorithm efficiency measure for different weights. Weights* (%) w1-w2 For all case studies described in Table 1 Simple Average Occupancy Behavior Fuzzy Average Algorithm efficiency measure*=w1 STD_BR+w2 NUM_ASBY

6 354 Intelligent Automation and Soft Computing occupancy algorithm is the second best performing algorithm followed by the average algorithm. The simple algorithms yields the lowest efficiency when the weight of total NUM_ASBY is equal or lower than the weight of the STD_BR. 4. EXPERIMENTAL SYSTEM 4.1 Description The implemented packaging system is composed of two articulated manipulators, a linear conveyor and an overhead camera. The package consists of three different part types, each part has a different size. A machine vision system identifies the arrival of a new part and its type. Each robot stocks the parts in a buffer, and prepares the package after all required parts have arrived (Figure 2). The buffer could house up to six parts. Camera Robot I conveyor Buffer Buffer Parts Robot II Figure 2. Experimental system layout Two algorithms were implemented: simple and fuzzy. Twelve experiments were conducted to compare the algorithms, while changing the parts arrival frequency and conveyor velocity. These parameters were varied in order to test the system s operation for different workloads and times available for decisionmaking. The system was fed with parts in three different frequencies: an average of 10 seconds between parts, an average of 15 seconds, and an average of 20 seconds. Conveyor velocities of 4 and 5 cm/sec were evaluated. Each experiment was conducted twice, once for each control algorithm. Each experiment included a batch of about 200 parts. 4.2 Results In the hardware implementation not all parts were taken due to a technical limitation in the part feeder mechanism. This was due to the fact that the parts had to be identified on-line. The vision system employed was not optimized for real-time performance, and therefore was the limiting factor in the system. In addition, the rate of parts not taken by any of the robots is high. Adding a cyclic conveyor to bring nontaken parts back to the system can solve this. Results were analyzed differently than in the simulation due to technical limitations. Two new parameters were defined: Number of parts taken (NPT) the number of parts picked by the robot's divided by the total number of part that entered the system; Normalized number of assemblies (NNA) the number of assemblies divided by the theoretical number of possible assemblies (the number of parts that entered the system divided by the number of parts in an assembly). Additionally, for verification, Utilization (the time the robot operated divided by the total experiment time) was

7 Distributed Multi-Robot Assembly/Packaging Algorithms 355 measured. Utilization, NPT, and NNA results should all point in the same direction and give compensating information. Utilization in the implemented system was relatively low (approximately 60%), due to the systems technical limitation. In all three parameters, Utilization, NPT, and NNA the simple algorithm yielded better results, 12%, 7%, and 5% difference respectively (Table III). Though similar in direction, in the simulation the average difference in the number of assemblies between the algorithms was much higher (12%) than the difference in the normalized number of assemblies (5%) [6]. This may be explained by the overall low utilization of the hardware experiment. Similar to the simulations workload was distributed much more evenly in the fuzzy algorithm. In the fuzzy algorithm the robots utilization was practically equal, similar to the simulation results. Table III. Average experimental results (in percentage). Simple Algorithm Fuzzy Algorithm Robot I Robot II System Robot I Robot II System NPT NNA Utilization As expected, increasing the parts arrival frequency increased average utilization from 46% in the lowest frequency to 59% in the highest frequency (13% improvement). However, it also increased the average number of parts not taken from 30% to 51% correspondingly [13]. An interesting trend observed was the change in the improvement rate [13]. Increasing the parts arrival frequency from 3 parts to 4 parts per minute improved utilization in the simple algorithm by 12%. Additional increase in parts arrival frequency improves the utilization by only 5%. However, in the fuzzy algorithm, increasing the parts arrival frequency from 3 parts to 4 parts per minute improved utilization by 7%. Additional increase in arrival part frequency improves utilization by 13%. As the mean time between parts arrival becomes shorter, the improvement rate for the simple algorithm decreases, while it increases in the fuzzy algorithm. To validate this trend statistically, additional experiments must be conducted. Increasing conveyor velocity did not reduce system performance showing that there was ample time for decision making for both algorithms in the current settings. 5. CONCLUSIONS This research investigated decentralized algorithms for multi-robot workcells and demonstrates its feasibility for an assembly/packaging task. The main advent in the presented algorithms is the robots abilities to deal with uncertainties in parts arrival. In these cases optimal sequencing procedures are not applicable, although future development could use them as a baseline for comparison in combination with a learning system that predicts parts arrival. Efficient and effective operation is enabled by independent operation of each robot. This efficiency increases as the system gets busier. The simple take-what-you-can algorithm yielded best performance regarding the total number of parts assembled both in simulation and in the experimental system. The fuzzy algorithm achieved best performance in terms of the efficiency measure defined. As the system gets busier this advantage is enhanced. The fuzzy algorithm was also effective in equalizing robots utilization. The hardware implementation indicated the need to define an additional performance measure system utilization. Future research is underway to validate the experiments with additional robots. ACKNOWLEDGMENTS This research was supported in part by the Paul Ivanier Center for Robotics and Production Management, Ben-Gurion University of the Negev.

8 356 Intelligent Automation and Soft Computing REFERENCES 1. A. Agah and A. G. Bekey, Efficiency Assessment of Performance of Decentralized Autonomous Multi-Robot System, Journal of Robotics and Mechatronics, 8(3), , R. C. Arkin, Behavior based robotics, MIT Press, MASS, F. Bard, E. Dar-El, and A. Shtub, An Analytic Framework for Sequencing Mixed Model Assembly Lines, Intl. Journal Production Research, 30(1), 35-48, S. Berman, M. A. Oliveira, Y. Edan, and M. Jamshidi, Hierarchical Fuzzy Behavior-Based Control of a Multi-Agent Robotic System, Proceedings of the 7 th Mediterranean Conference on Control and Automation (MED99), R. W. Bernnan, Performance comparison and analysis of reactive and planning-based control architecture for manufacturing, Robotics and Computer integrated Manufacturing, 16, , E. Boteach, Algorithms for multi-robot assembly/packaging systems. M.Sc thesis, Dept. of Industrial Engineering and Management, Ben-Gurion University of the Negev, Beer Sheva, Israel, R. A. Brooks, A robust layered control system for a mobile robot. IEEE Trans. on Robotics and Automation 2(1), 14-23, P. H. Brill and M. Mandelbaum, Measurement of adaptivity and flexibility in production system European Journal of Operational Research, 49, , J. Connell, A behavior-based arm controller, IEEE Trans. on Robotics and Automation, 5(6), , E. Gat, R. Desai, R. Ivlev, J. Loch, and D. Miller, Behaviour control for robotic exploration of planetary surfaces. IEEE Trans. on Robotics and Automation 10(4), , P. Y. Glorennec, Coordination between autonomous robots. Intl. J. of Approximate Reasoning 17(4), , M. Jamshidi, Large scale systems: modeling control and fuzzy logic, Prentice Hall, NJ E. Mendelson and O. Nayer, Implementation of Multi-Robots Behavior Based With Fuzzy Logic Packaging System. B.Sc. Final Project, Department of Industrial Engineering, Ben Gurion University of the Negev, Beer-Sheva 84105, S. Y. Nof and D. Hanna, Operational characteristics of multi-robot systems with cooperation. Intl. Journal Production Research 27(3), , L. E. Parker, Alliance: An Architecture for fault tolerant multirobot cooperation. IEEE Trans. on Robotics and Automation 14(2), , V. Rajan and S. Y. Nof, A game theoretic approach for cooperation control in multi-machine workstations. Research Memorandum No School of Industrial Engineering, Purdue University, R. S. Sharma, S. M. LaValle, and S. Hutchinson, Optimizing robot motion strategies for assembly with stochastic models of the assembly process. IEEE Trans. on Robotics and Automation 12(2), , T. Shibata, K. Ohkawa, and K. Tanie, Spontaneous behaviour of robots for cooperation - emotionally intelligent robot system. Proceedings of the 1996 IEEE Intl. Conf. on Robotics and Automation, , D. Shirley and J. Mateijevic, Mars pathfinder microver. Autonomous Robots 2, , C. P. Tung and A. Kak, Integrating sensing, task planning and execution for robotic assembly. IEEE Trans. on Robotics and Automation 12(2), , S. T. Yee and J. A. Ventura, A petri net model to determine optimal assembly sequences with assembly operation constraints. Journal of Manufacturing Systems 18(3), , 1999.

9 Distributed Multi-Robot Assembly/Packaging Algorithms 357 ABOUT THE AUTHORS Y. Edan received a B.Sc. in Computer Engineering, Technion; M.Sc. in Agricultural Engineering, Technion and PhD in Engineering, Purdue University. She has performed research in robotics, sensors, system architectures, simulation, and decision-making systems. In addition, she has made major contributions in the introduction and application of intelligent automation and robotic systems to the field of agriculture with several patents. She is currently the Chair of the Dept. of Industrial Engineering and Management at Ben-Gurion University of the Negev, and was the Chair of the Paul Ivanier Center for Robotics and Production Management at Ben-Gurion University. Prof. Edan is a member of the IEEE Robotics and Automation Society, and ASAE, Society for Engineering in Agricultural, Food and Biological Systems. She has been the Chair of the Flexible Automation and Robotics/Mechatronics and BioRobotics committees, ASAE. Dr. Edan is active on several editorial boards (Intl. Journal of Industrial Engineering - Applications and Practice; IIE Transactions on Operations Engineering) and has been an Associate Editor of Transactions of the ASAE. S. Berman received a B.Sc. in Electrical and Computer Engineering, Technion; M.Sc in Electrical and Computer Engineering, and Ph.D. in Industrial Engineering, both from Ben-Gurion University of the Negev. In she was a visiting scholar at the ACE Center, the University of New Mexico, USA. She currently is a Post doctoral fellow at the Department of Computer Science and Applied Mathematics at the Weizmann Institute of Science. Her research interests include robotics, computer integrated manufacturing, and soft computing. Photo Not Available E. Boteach received a B.Sc. in Mechanical Engineering, and M.Sc. in Industrial Engineering and Management, both from Ben-Gurion University of the Negev. He currently is an engineer in a software company in Israel. E. Mendelson received a B.Sc. in Industrial Engineering and Management, Ben- Gurion University of the Negev and this work is part of her M.Sc. thesis in the Dept. of Industrial Engineering and Management, Ben-Gurion University of the Negev. She is a PhD candidate at the Interdisciplinary Center for Neural Computations at the Hebrew University in Jerusalem. Her research interests include robotics and industrial engineering.