Learn to Grasp Utilizing Anthropomorphic Fingertips together with a Vision Sensor

Transcription

1 Learn to Gra Utilizing Anthroomorhic Fingerti together with a Viion Senor Yaunori Tada, Koh Hooda, and Minoru Aada Adative Machine Sytem, Graduate School of Engineering HANDAI Frontier Reearch Center Oaka Univerity 2-, Yamadaoka, Suita, Oaka , Jaan tada@er.am.eng.oaka-u.ac.j, {hooda, aada}@am.eng.oaka-u.ac.j Abtract A robot hould have oftne and many enor to maniulate an object dexterouly and to adat variou environment. However, many exiting cheme that a deigner calibrate the enor outut to the world coordinate frame are difficult to adat for uch the robot. Thi aer rooe a learning mechanim for a robot hand which conit of anthroomorhic fingerti. The enor for the fingerti i difficult to calibrate becaue the enor recetor are embedded randomly in the oft material. The effectivene of the rooed mechanim i demontrated by an exeriment that the robot ick u an unknown weight object. Index Term graing, object maniulation, ditributed enor, oft finger, anthroomorhic finger I. INTRODUCTION Many exiting control cheme for the robot hand are deigned with reect to the world coordinate frame, and outut of the enor are calibrated with the frame by the deigner [], [2], [3]. Thee robot achieve a given tak by uing model of the robot and the object decribed with the frame. Therefore, the robot and the object have to be imle enough to be eaily modeled by imle calibration. For thi uroe, the deigner ha to lace the enor in the exact oition and orientation. However, the robot uing the cheme can not adat to unknown environment which the deigner did not uoe. Furthermore, the robot uing the cheme i affected by calibration error and diturbance. If the enor lacement for the robot i fully controlled, it may not be able to ene the information that i eential for the tak but that the deigner ha not exected. Therefore, dexterity of the exiting robot i not ufficient becaue of lack of ening ability. One way to avoid uch deigner bia and limitation of ening i to embed a many recetor a oible into the robot randomly. Our reviou work [4] how that the oft fingerti which contain many tactile recetor ha high ening ability. However, the robot which ha oftne and many enor i difficult to calibrate the enor and to build a controller. Therefore, the tak and the enor outut mut be rereented in a enor ace which i defined by the robot elf. In [5], the outut of the viion and force enor were rereented in the enor ace of the robot, and the robot achieved maniulating an object. Hooda et al. [6] roduced an anthroomorhic fingerti whoe tactile recetor are randomly ditributed in and on the oft material. Furthermore, a neural network for the rereentation of liage utilizing the anthroomorhic fingerti and a viion enor wa rooed. The context of the given tak i, however, not conidered in the mechanim. Additionally, Hakozaki el al. [7] rooed artificial kin in which many tactile recetor are embedded randomly, what they call telemetric kin. However, they rooed only the tructure and did not rooe the control architecture to utilize it. Yamada et al. [8] rooed a oft finger whoe tactile recetor are embedded regularly, and which can gra the unknown weight and a friction coefficient object by detecting the inciient liage. However, the deigner decided the mechanim for detecting the inciient liage and for controlling the graing force. Thi aer rooe a learning mechanim to utilize the anthroomorhic fingerti through a tak context, that i, a graing tak, together with a viion enor. The mechanim find the correlation between the outut of the viion and the tactile enor, and realize graing. Exerimental reult to ick u an object are hown to demontrate the effectivene of the rooed mechanim. The remainder of thi aer i organized a follow. Firt, the deign of the anthroomorhic oft fingerti i introduced. The fingerti conit of two ilicone rubber layer of different hardne containing two kind of randomly ditributed recetor. Then, the learning mechanim for graing i rooed. It learn the correlation among the outut of the enor and a motor command by Hebbian learning. Finally, exerimental reult are hown to validate caability of the rooed mechanim. II. ANTHROPOMORPHIC SOFT FINGERTIP Fig. how a cro ectional view of the develoed anthroomorhic oft fingerti. The fingerti ha two layer made of the different hardne of ilicone rubber: the outer layer i harder than the inner layer. A rod i inerted at the

2 recetor: train gauge and PVDF film an outer layer an inner layer a metal rod Fig.. A cro ectional view of the develoed anthroomorhic fingerti center of the fingerti to lay a role of a bone. Strain gauge and PVDF (olyvinylidene fluoride) film that ene train and velocity of train are embedded in both layer a tactile recetor. A develoed oft fingerti i hown in Fig. 2. It diameter and length are 25 mm and 55 mm, reectively. The fingerti ha twenty-four recetor; ix train gauge and ix PVDF film are embedded in the inner layer, and alo the ame number of the train gauge and the PVDF film are embedded in the outer layer. Since the recetor are embedded randomly, the ame kind of recetor embedded in the different oition may ene different kind of tactile information. A train gauge embedded near the urface of the kin i exected to ene the local tatic train between the kin and the object urface wherea the train gauge embedded near the rod i exected to ene the total force exerted on the finger and i exected to be inenitive to the local texture of the object. A PVDF film i exected to ene the train velocity, which mean that it i more enitive to the tranient or mall train (or tick li) than the train gauge. The ilicone rubber between two PVDF film i exected to act like a low-a filter; therefore, the difference between the ignal i exected to rereent the local tick li henomena. Additionally, our reviou work [4] how that the fingerti ha high ening ability by dicriminating between the object. III. LEARNING MECHANISM FOR GRASPING The calibration for the anthroomorhic fingerti i difficult becaue the fingerti i made of oft material and contain randomly ditributed tactile recetor. Therefore, the robot ha to learn the meaning of the enor outut intead of the calibration. In thi aer, the robot hand with a viion enor iterate icking u an object, and learn the correlation among the tactile enor, the viion enor, and a motor command. From thi learning, the robot acquire behavior for icking u the object by minimum graing force. Exerimental Fig. 2. A hoto of the anthroomorhic oft fingerti condition for the learning mechanim are a follow: ) weight and a friction coefficient of the object are unknown, 2) the graing force i increaed or decreaed dicretely by a ign of the controller outut, and 3) the viion enor oberve motion between the object and the robot, then it outut three tate: icking u ucce, toing the robot hand, and icking u failure. Fig. 3 how a block diagram of a control ytem for the robot. Detail of neural network which conit of a enor network and a motor network are hown in Fig. 4. The enor network learn the correlation between the outut of the tactile enor and the viion enor, and outut a tate of the graing object. The motor network convert the outut of the enor network to the motor command. Detail of the enor network in Fig.4 i a follow. The enor network learn the correlation between the outut of the viion and the tactile enor by Hebbian learning. The outut of each train gauge i normalized by the maximum value over time and i given to the train gauge n. n range from to. Furthermore, the outut of each PVDF film i alo normalized by the maximum value over time, and the value of or are given to the PVDF film k if the normalized value i over or under a threhold, reectively. The value of,, and are given to the viion v and denote icking u failure, toing the robot hand, and icking u ucce, reectively. The outut of the enor network i calculated by following equation. S = k S = n k w k + v () n w n + S (2)

3 deired value tactile inut -~ train gauge, PVDF film + - Fig. 3. enor network 2 n 2 k w w w 2 w n w w w 2 w k Fig. 4. motor command motor network robot hand enor network A block diagram of robot hand controller train-viion v S S viion -,, - + PVDF-viion d deired value motor network enor-motor w v m motor -, A neural network for graing behavior graing object where, w k denote the connection weight between the PVDF film and the viion, and wn denote the connection weight between the train gauge and the PVDF viion. The initial value of w k and w n are et to zero. The connection weight of the enor network are udated by following equation. S a = a nwn n (3) w k = α k v (4) w n = β n S ( S a ) (5) where, α and β denote learning rate. By (4), the PVDF viion learn the correlation between the outut of the PVDF film and the viion enor when the fingerti li on the object becaue the PVDF film reond to velocity of train. In (3), a n i the value of the train gauge at intant of icking u the object, and w n i the current connection weight. The connection weight w n i udated that S a aroximate in (5). In the equation, the contant value of i the ame value a the PVDF viion S when the S m robot achieve icking u the object. Therefore, the train viion learn minimum graing force for icking u the object by (3) and (5). Thu, the tactile enor lay the ame role a the viion enor by uing thi neural network. Additionally, after the learning, the enor network outut the value of when the robot achieve icking u the object. When the graing force i inufficient, the network outut le than. The motor network in Fig.4 learn the correlation between the outut of the enor network S and the motor command m by Hebbian learning. Thu, the network outut the motor command for controlling the graing force by ubtracting S from the deired value d. The deired value d i the ame value a the train-viion S when the robot achieve icking u the object: the value of d equal. The motor command m denote increaing or decreaing the graing force, and the value of or are given to m. Thi value i given at learning hae only, and the value equal at after learning. An outut of thi network S m i calculated by a following equation. S m = (d S )w v + m (6) where, w v denote the connection weight between the outut of the enor network and the motor. The initial value of w v i et to zero. The connection weight of the motor network i udated by a following equation. w v = γ S m (7) where, γ denote a learning rate. Conequently, the robot acquire graing behavior that adat to liage by uing the enor network and the motor network. IV. EXPERIMENT To demontrate caability of the rooed learning mechanim, an exeriment that the robot ick u an object i hown. The object i a latic cu that it weight and a friction coefficient are unknown. The robot autonomouly acquire a motor command and the minimum graing force to ick u the object. Furthermore, the exeriment how that the robot can deal with the heavier object by adating to liage. A. Exerimental equiment Fig. 5 and 6 how an exerimental equiment. The develoed fingerti are mounted at the end of a robot hand that ha 4 DOF by two finger (Fig. 5). However, the hand i ued for DOF grier becaue the exeriment demontrate controlling the graing force. Additionally, becaue of limit of the number of A/D converter channel, the tactile recetor in one ide fingerti are ued for the tactile enor. The outut of the train gauge and the PVDF film are amlified and fed to a hot comuter via an A/D converter. A robot hand ytem contain the robot hand, a robot arm, and a video

4 Outut of viion enor - touching liing icking u ucce end of exeriment 6 Outut of train gauge (V) Fig. 5. The anthroomorhic fingerti are mounted at end of robot finger. Outut of a PVDF film (V) Time () Fig. 7. An examle of enor outut when the fingerti ick u the object Fig. 6. A robot ytem conit of a robot arm, a robot hand with anthroomorhic fingerti, and a video camera. camera (Fig. 6). The object i icked u by the robot arm, and the camera oberve the motion between the object and the arm. An examle of the outut of the viion and the tactile enor when the robot hand ick u the object i hown in Fig. 7. Thee curve in the figure are art of the outut of the tactile enor that reond largely to a contact. In the examle exeriment, the graing force and the robot arm are controlled by a deigner. The exerimental rocedure conit of two hae. Firt, the robot hand i moved to a lower oition of the object. In thi hae, the fingerti do not touch the object. Then, the object i icked u by increaing the graing force and moving the robot hand to the uer oition. In Fig. 7, the fingerti doe not touch the object at the beginning of the exeriment, and the robot hand tart moving to the uer oition when the fingerti touche the object. The fingerti touche the object at about.9. Until about.8, the fingerti li on the urface of the object becaue the graing force i inufficient. Then, the graing force become ufficient, and the object i icked u until about 5.6. Thi figure how three characteritic of the enor: ) the viion enor outut tate of the graing object, 2) the train gauge reond to graing force, and 3) the PVDF film reond to tranition of train uch a an intant of contact or liage. B. Learning rocedure The develoed fingerti ha twelve train gauge and the ame number of the PVDF film. However, ome uitable recetor to achieve the tak are elected by the deigner: the train gauge that maximum outut i more than.5 V are elected, and the PVDF film #3 which largely reond to liage i elected. The learning rocedure conit of three hae. ) The robot hand i located at a bottom of the object. At thi time, the fingerti do not touch the object. 2) The graing force i increaed, and the robot hand move to the uer oition. In thi learning hae, the graing force i increaed regardle of the outut of the motor network S m. 3) The connection weight are udated by the outut of

5 Outut of the viion enor Normalized outut of train gauge Normalized outut of a PVDF film Connection weight at the train-viion (olid line) trial w 3 Time () w 5 w Connection weight at the PVDF-viion (broken line) Fig. 8. Time coure of the enor outut and the connection weight at learning hae the tactile enor, the viion enor, and the motor command. 4) The above three hae are iterated. Additionally, the robot arm i controlled by the deigner. The connection weight w n, w k, and wv are reinforced by thi rocedure, then the neural network acquire the graing force control to ick u the object. At the hae 2), the graing force i increaed regardle of the outut of the motor network S m. The reaon i that the outut of the motor network equal at the beginning of the learning becaue the initial connection weight equal, then the graing force doe not increae. Therefore, to begin udating the connection weight, the motor command excet the value of i needed. Additionally, becaue of the linear tructure, the neural network can outut the motor command for decreaing the graing force even if the network learned a ituation of increaing the graing force only. Fig. 8 how time coure of the enor outut and ome connection weight at the learning hae. The robot doe not ick u the object while the viion enor outut the negative value. In thi tate, the fingerti li on the urface of the object becaue of the inufficient graing force. When the graing force become ufficient, the robot ick u the object, and the viion enor outut the oitive value. Therefore, the train viion learn minimum graing Outut of viual enor Normalized outut of train gauge Normalized outut of a PVDF film Connection weight at the train-viion ouring water into the cu Time () Fig. 9. Time coure of the enor outut and the connection weight at after learning force by (5). On the other hand, the large ignal of the PVDF film occur when the viion enor outut the negative value, and the PVDF film doe not outut the large ignal when the viion enor outut the oitive value. Therefore, the PVDF viion learn liage by (4). Becaue of linear tructure, the network can not learn the excluive tate that the robot fail and ucceed in icking u the object. Therefore, in the bottom of Fig. 8, the connection weight w 3 and w 5 at firt trial are not table by the outut of the viion enor. However, the connection weight are exected to converge becaue the learning at the failing in icking u the object i almot ignored. The reaon why the learning i almot ignored are a follow: ) the number of learning at ucceeding in icking u the object i more than failing in icking u one, 2) the udating value w n in (5) at failing in icking u the object i mall becaue the outut of the train gauge i mall by the inufficient graing force, and 3) the udating value w n become mall te by te becaue ( S a ) in (5) i aroximated by udating the connection weight w n. Therefore, the connection weight of the train viion w n are converged after ome trial. In the figure, the connection weight are converged by five trial. w 5 w 3

6 C. Picking u the object at after learning An exeriment demontrate that the robot hand uing the learned neural network can ick u a cu a the object by minimum graing force, and that the robot can deal with the heavier object. The rocedure of thi exeriment i the ame a the rocedure in the learning hae excluding that the fingerti touche the object at the beginning of the exeriment. Furthermore, the viion enor v equal at any time to demontrate that the tactile enor lay a role of the viion enor. However, the outut of the viion enor i recorded for analyzing exerimental data. Additionally, the motor command m alo equal. Note that the connection weight w n are udated by the outut of the train gauge and the PVDF film even if the viion equal. Fig. 9 how time coure of the outut of the viion enor, the train gauge, the PVDF film, and the connection weight w n. From thi figure, the outut of each train gauge until 2 i maller than the learning hae outut in Fig. 8. In other word, the robot gra the object by mall force. Then, the weight of the object become heavy by ouring water into the cu at 2 to 4. The PVDF film outut large ignal at thi time becaue of imact of ouring water. After ouring water, the liage between the cu and the fingerti occur becaue of the inufficient graing force. The PVDF film reond to the liage, and the PVDF viion outut the tate of the icking u failure even if the viion equal. It mean that the PVDF film lay a role of the viion enor. Then, the connection weight w n are udated by only the outut of the tactile enor. Conequently, the liage between the fingerti and the object i toed by increaing the graing force. Thi reult how that the robot can adat to heavier weight by the rooed mechanim. V. DISCUSSION AND FUTURE WORKS In thi aer, the learning mechanim for the fingerti which ha the randomly ditributed recetor wa rooed. The learning mechanim wa to find the correlation among the outut of the enor and the motor command, and to realize the graing. The exeriment to demontrate the ability of the mechanim howed that the robot can ick u the unknown weight object through learning, and can deal with the heavier object. The tactile enor of the anthroomorhic fingerti i imilar to human tactile organ which ene variou contact condition. The recetor of the tactile organ can ene different roertie becaue of the dynamic of the body where the recetor i embedded even if the recetor have imilar reone characteritic. Thee recetor can ene contact force, liage, or vibration that are meaningful contact information to achieve dexterou maniulation [9], []. Therefore, we exect that the robot hand can maniulate the object dexterouly by utilizing the anthroomorhic fingerti even if the exerimental reult in thi aer only how the control of graing force. Although the anthroomorhic fingerti ha twenty-four tactile recetor, in the exeriment, ome uitable recetor to achieve the tak are elected by the deigner. The autonomou robot ha to elect uitable recetor by itelf. The exeriment how that the robot control graing force only. The robot ha to achieve comlex tak to demontrate the caability of the fingerti. Additionally, the fingerti can not dicriminate between the contact and the liage becaue the PVDF film ene the velocity of train. Thee are future work. ACKNOWLEDGMENT Thi tudy wa erformed through the Advanced and Innovational Reearch Program in Life Science, and Grantin-Aid for Scientific Reearch (B) from the Minitry of Education, Culture, Sort, Science and Technology of the Jaanee Government. REFERENCES [] T. Matuoka, T. Haegawa, and K. Honda, A dexterou maniulation ytem with error detection and recovery by a multi-fingered robotic hand, In Proceeding of the 999 IEEE/RSJ International Conference on Intelligent Robot and Sytem, volume, , 999. [2] Y. Yokokohji, M. Sakamoto, and T. Yohikawa, Viion-aided object maniulation by a multifingered hand with oft fingerti, In Proceeding of the 999 IEEE International Conference on Robotic and Automation, , 999. [3] T. Watanabe, K. Harada, Z. Jiang, and T. Yohikawa, Object Maniulation under Hybrid Active/Paive Cloure, In Proceeding of the 25 IEEE Internal Conference on Robotic and Automation, , 25. [4] Yaunori Tada, Koh Hooda, and Minoru Aada, Sening Ability of Anthroomorhic Fingerti with Multi-Modal Senor, In Proceeding of the 8th Conference on Intelligent Autonomou Sytem,. 5 2,24. [5] Koh Hooda, Takuya Hiano, and Minoru Aada, Senor Deendent Tak Definition: Object Maniulation by Finger with Uncalibrated Viion, In Proceeding of Intelligent Autonomou Sytem 6, , 2. [6] Koh Hooda, Yaunori Tada, and Minoru Aada, Internal Rereentation of Sli for a Soft Finger with Viion and Tactile Senor, In Proceeding of the 22 IEEE/RSJ International Conference on Intelligent Robot and Sytem,. 5, 22. [7] Mituhiro Hakozaki, Katuhiko Nakamura, and Hiroyuki Shinoda, Telemetric artificial kin for oft robot, In Proceeding of TRANS- DUCERS 99, , 999. [8] Daiuke, Yamada, Takahi Maeno, and Yoji Yamada, Artificial Finger Skin having Ridge and Ditributed Tactile Senor ued for Gra Force Control, Journal of Robotic and Mechatronic, Vol. 4, no. 2,. 4 46, 22. [9] R. S. Johanon, U. Landtrom, and R. Lundtrom, Reone of mechanorecetive afferent unit in the glabrou kin of the human hand to inuoidal dilacement, Brain Reearch, no. 244,. 7 25, 982. [] R. S. Johanon and A. B. Vallbo, Tactile enory coding in the glabrou kin of the human hand, Trend in Neurocience, no. 6, , 983.