NAIST Openhand M2S: A versatile two-finger gripper adapted for pulling and tucking textiles

Transcription

1 2017 First IEEE International Conference on Robotic Computing NAIST Openhand M2S: A versatile two-finger gripper adapted for pulling and tucking textiles Felix von Drigalski, Daiki Yoshioka, Wataru Yamazaki, Sung-Gwi Cho, Marcus Gall, Pedro Miguel Uriguen Eljuri, Viktor Hoerig, Ming Ding, Jun Takamatsu and Tsukasa Ogasawara Robotics Laboratory, Graduate School of Information Science Nara Institute of Science and Technology (NAIST) Takayama, Ikoma, Nara , Japan (felix.von drigalski.fp6, yoshioka.daiki.xr0, wataru.yamazaki.yo6, cho.sungi.cg3, marcus.gall.lw3, pedro.uriguen.pl3, viktor.hoerig.uw4, ding, j-taka, ogasawar)@is.naist.jp Jessica Beltran Laboratory of Safety Intelligence System Department of Mech. Sc. and Eng., Nagoya University Furo, Chikusa, Nagoya, Aichi , Japan jessica.g.beltran@ieee.org Abstract When manipulating textiles and thin objects, one challenging task is to push (tuck) of textiles into small openings. Further, for a number of manufacturing tasks it poses an obstacle to automation. However, current robot grippers are almost exclusively designed for grasping objects, or imitate the human hand with very fine mechanisms that break easily, so that they cannot withstand the loads required by the pushing task. Further, as textiles are very thin, even grippers with pressure pads cannot easily confirm if a textile was grasped successfully or if the gripper is empty. In this research, we present a gripper design that can detect the successful grasping of thin objects via active perception, can sustain significant pushing loads in order to perform tucking tasks and can perform dexterous grasping and in-hand manipulation. The gripper is open-source and can be 3D printed. We demonstrate the gripper s performance experimentally, its precision when controlling its grasp force, and the maximum grasping force. Keywords-robot grippers; textile manipulation; open-source manipulators; I. INTRODUCTION As textile manipulation is required for numerous tasks in the household (e.g. laundry, bed making, table setting) as well as the workplace (e.g. car seat assembly, furniture manufacture, dry cleaning facilities), there is a high demand for versatile robots that can manipulate textiles and clothing articles as dexterously as rigid objects. One of the tasks that is still not fully automated is the pushing (tucking) of flat deformable objects into small openings. To advance in the automation of textile manipulation tasks, a gripper is desirable that is able to perform both sensitive precision grasping as well as tucking task. However, most robotic manipulators are designed only for grasping, so that they cannot support an axial load on the tip of the gripper. Additionally, many grippers cannot detect if a textile has been grasped, due to the thinness Figure 1. First prototype of the proposed gripper with force sensors Table I COMPARISON OF ROBOT HANDS AND THEIR FEATURES. Hand name Barrett Hand Robotiq 2-finger Robotiq 3-finger Schunk 5-finger Schunk 3-finger Sake EZGripper Our proposal Tactile feedback Axial load Flat fingers Relative sensor motion 2 and deformability of textiles. Lastly, many grippers are not compliant, so that it is hard to realize a grasp with very little force, such that silk would slide through the gripper s fingers. An overview of common grippers and their features is listed in Table I /17 $ IEEE DOI /IRC

2 Figure 2. Human pulling a textile taut with two hands and light grasp force during bed making Outside of a few exceptions [1], robotic grippers used in textile manipulation research are most often 1-DOF 2-finger grippers and do not make use of sliding points of contact with the object. The level of research interest in the field of home laundry automation and cloth folding is perhaps best represented by the headway made by Pieter Abbeel s group [2], [3], but contributions are numerous [4] [6]. We based our design on the Yale OpenHand M2 gripper [7], which is one in a series of open-source manipulators proposed by Dollar et al. [8]. While other open-source gripper designs are publicly available [9], [10], none feature a rigid thumb that can sustain axial loads. Ward-Cherrier et al. have added an optical tactile sensor to the M2 gripper [11]. However, their sensor is mounted on the thumb, which increases its size and prohibits pushing into tight spaces or with high forces. Lastly, Kaboli et al. [12] use the same sensors as in this work, mounted on a Robotiq 3-finger gripper to detect slip in deformable objects. Their results support our assumption that the sensors yield sufficient information for textile manipulation. We propose a gripper that can perform both tucking tasks as well as sensitive grasping manipulations for textiles. Our goal is to implement previously unautomated textile manipulation and manufacturing tasks, such as stretching out a cloth or clothing article with two arms, pulling covers over articles, and tucking textiles into tight openings. A prototype can be seen in Fig. 1. Our contribution consists of the combination of the improved mechanical design of the gripper, the source code, the force sensors and the approach to detect the successful grasp of a textile by generating a relative motion between the sensors. To the best of our knowledge, evaluating grasp success of thin objects and textiles only by friction response 2 Relative sensor motion is the ability to move tactile sensors relative to each other when in contact. Figure 3. Representation of a textile being pulled taut while sliding through the gripper s fingers Figure 4. KUKA LBR iiwa with a prototype of the proposed gripper pulling a bed sheet over a mattress is novel. We evaluate the feasibility of the design experimentally. The remainder of this paper is structured as follows: In section 2, we explain the mechanical design of the gripper and our design approach. In section 3, we show the performance of the gripper in exemplary tasks. Lastly, we discuss the results and explore directions for future work. II. DESIGN &HARDWARE A. Approach We consider textile manipulation to consist of elementary manipulations such as grasping, sliding and tucking. Grasping creates a temporary, fixed connection between one or more points and the grasping surface of the grippers (e. g. the fingertips). Sliding is considered to occur when relative movement between the object and the grasping surface takes place, 118

3 Figure 5. Human tucking a bed sheet in between mattress and frame during bed making Figure 6. KUKA LBR iiwa with a prototype of the proposed gripper tucking a bedsheet under a mattress but the grasp is not necessarily lost after the operation, as shown in Fig. 2 and Fig. 3. Tucking is the insertion of part of a textile into a small opening or crack by pushing onto the textile, as shown in Fig. 5, Fig. 6 and Fig. 7. We enable all three of these elementary manipulations with our gripper design. In order to automate tasks involving tucking, we saw a need for a protruding part with a small profile that can sustain axial loads and push with its tip, so as to apply pressure on a small area. Further, to be able to use the gripper in our research we required a design to which we can add sensors freely. Consequently, we based the design on the Yale OpenHand M2 gripper [7], an under-actuated gripper with two degrees of freedom on one finger, and a thumb without joints. We use the thumb to tuck textiles into cracks, and added two 3D force sensors to the hand s fingertips. This allows not only the compliant grasping of objects and textiles, but also to grasp with small amounts of force let a textile slide between the gripper s fingers. We also use the sensors to evaluate the grasp success of textiles by using a relative motion. In the spirit of open-source design, both our code and design are freely available online. 3 B. Hardware The gripper shares most of the basic characteristics with the M2 gripper, such as one rigid thumb, a 2-DOF finger with an agonist and antagonist tendon arrangement, and a base with actuating motors. The agonist motor closes the second finger joint and results in an underactuated grasp, while the antagonist motor results in a fully actuated grasp where the second finger joint does not close. The main additions to the original design are the two force sensors 3 Figure 7. Representation of the tucking manipulation. The arrow represents a gripper pushing into the space between the two objects and the bearings in the finger joints, and the omission of any rubber surfaces and hybrid deposition techniques, which makes it simpler to print and assemble than the original design. Our proposed design consists of only one 3D printed material and commercially available parts. We added ball bearings with an outer diameter of 5 mm to each finger joint, in order to reduce friction and thus increase the maximum grasping force. With decreased friction, the force of the springs retracting the finger can be significantly reduced, which lowers the load on the motors that have to counteract the spring force during operation. While the bearings can be omitted, they make a significant difference when using less costly servo motors. In total, the gripper contains: 2 OptoForce 3D force sensors (OMD-20-SE-40N) 1 Arduino Uno microcontroller 2 HS-5585MH servo motors 4 3D printed base & finger links Miscellaneous bearings, pins, pulleys, cables The hand is mounted on a KUKA LBR iiwa 14 R820 robot arm, which is equipped with torque sensors. We use the robot s torque sensors to control pushing tasks, and the OptoForce sensors for grasping and manipulation. One problem that can arise after pushing a textile into an opening is the textile being pulled out along with the 4 A base design for Dynamixel MX28-AT motors is also available. 119

4 finger. This occurs when the friction between the robot finger and the textile is higher than between the textile and the surrounding objects. To reduce the chance of textile pullout after tucking, we aim to reduce the finger s friction by omitting the rubber surface of the original design and by limiting the height of the force sensor protruding from the finger surface. While the force sensors are capable of precision grasps, if an object is in a power grasp, they do not report all the force acting on it, as it is in contact with other parts of the hand than the sensors. Fully actuated grasps may also result in a grasp where the force sensors are not in contact with the object. We consider these limitations minor, as our main focus is the manipulation of textiles. We note that as long as the force sensors are in contact with the object, slip can still be detected. Aside from the addition of the bearings and the sensors, the assembly of the gripper can be completed by following the tutorial for the original design of the M2 gripper. If heavy objects are to be lifted, adding rubber to the grasping areas of the finger and thumb will help increase friction. Figure 8. Grasp force at different motor positions. Motor 1: Antagonist tendon. Motor 2: Agonist tendon. III. EXPERIMENTS We experimentally evaluate the potential and functions of the hand with three experiments. A. Maximum force in different positions In this experiment, we sweep a grid of motor positions and record the force reported by the sensors, as shown in Fig. 8. This reveals the maximum grasping force in different configurations and grasping angles, and can also be used for the creation of a model-based controller. The maximum grasp force at the sensors is 4700 mn, as shown in the graph. We note that the effective grasp force can be different for larger objects such as cylinders, and it may be only partially recorded by the sensors. Sensors were run at 30 Hz with a low-pass filter during this experiment, as well as the following one. B. Force control In this experiment, we use a simple controller to achieve a target grasp force for the hand and to evaluate the effective precision of the hand. The controller is simple: the motor step of the antagonist is increased by one when the force is below the dead zone and vice versa, while the agonist tendon is still. As this utilizes the minimum discrete signal that a controller can use, it demonstrates the maximum precision of the setup. While a properly adjusted PID or model-based controller would be faster to reach the target force and may exhibit lower vibration, it will not be more precise than shown in this experiment unless the actuator is changed. The controller was updated once every 150 ms. The data shows that at worst, the grasp force precision is about ±0.25 N. An example graph can be seen in Fig. 9. Figure 9. C. Textile perception Total force seen by the force sensors when targeting 2 N In this experiment, we define two sets of motor angles which generate a back-and-forth motion that rubs the force sensors on each other. In between the force sensors, we place three different materials. See Fig. 10 for an illustration and i3gso for a video of the experiment. The motion creates vibrations as well as tangential forces, which can be used to distinguish if a textile is grasped between the sensors. The sensors were run at 1000 Hz with no filter during this experiment. Fig. 11 shows the lateral force seen by one of the force sensors during this motion for four different cases: without any material between the sensors, with one layer of cloth, with two layers of cloth, and finally with four layers of plastic (e.g. a folded garbage bag). The gripper was mounted horizontally and the material placed freely onto the thumb s sensor, as shown in Fig. 12. As the friction between the two 120

5 ms and 3100 ms in Fig. 11, when the finger moves in the opposite direction and the lateral force changes direction: the absolute difference between the force just before the movement and the peak that followed was 19% and 80% lower when the cloth and plastic were grasped. Figure 12. Experiment setups for different materials (left to right: empty, textile, plastic) Figure 10. Picture of the motion rubbing the sensors together. The finger on the left moves back and forth between the two positions pictured, causing tangential forces to occur due to friction Figure 11. Lateral force when rubbing sensors on each other while grasping different materials sensors is significantly higher than between a sensor and most other thin objects and textiles, the lateral force differs between objects. This is most clearly seen at around 2050 IV. DISCUSSION One limitation of the force sensor data is that while the force recorded is the lateral force as seen by the force sensor, it is not the tangential contact force, as the point of contact is generally not at the tip of the sensor. This is also the reason why the graph is not symmetric around zero: the force sensor is inclined and the point of contact changes during the movement. In order to convert the lateral force to the tangential contact force, one would need to know the point of contact between the sensors, but as the gripper is underactuated and without joint encoders, the position of the sensors is unknown. Nonetheless, the data implies that detecting a grasped textile from the force response profile generated by the rubbing motion is possible. Further, the difference between the behavior of one and two layers of textile is notable and implies significant tribological interaction between the materials. Evaluating the features of the force signal and vibrations, investigating their limits for slip detection and material recognition and training an appropriate classifier for these use cases is part of our future work. During the force control experiment, in some configurations especially if the finger s sensor is closer to the base than the thumb s sensor increasing the motor step of the antagonist does not press the sensors against each other, but slides them off on each other, which can complicate the control slightly. For the tucking task, we plan to use the robot s torque sensors to guide the tool during the tucking task, and disregard the force sensor s data. However, if a strong tangential force is detected on the thumb s sensor during tucking, it would be likely that the textile has been pulled out along with the gripper. Consequently, while the sensor data may be hard to interpret during tucking, it may support the recognition of failures to some degree. Another direction we plan to investigate is a redesign of the finger s tip so that the hand can pick up textiles from a flat surface using only one motor and without movement of the robot arm. This would simplify many grasp planning operations with textiles significantly, and make the gripper more useful. V. CONCLUSION We propose a versatile, robust gripper with tactile feedback that can grasp, slide and tuck textiles, and detect via active perception if a textile has been successfully grasped. We demonstrate that the grasping force at its fingertips can be controlled within 0.3 N or less, and a maximum grasp force 121

6 of 4.5 N can be delivered in even the least favorable configuration. It offers an inexpensive solution for tactile manipulation and can be used to investigate a variety of dexterous manipulation tasks, particularly the handling of textiles and clothing articles. The design is open-source and available on our website: Future work will focus on investigating the possibility of using the sensors for material recognition and simplifying the action of picking up a flat object with the gripper. REFERENCES [1] C. Elbrechter, R. Haschke, and H. Ritter, Folding paper with anthropomorphic robot hands using real-time physics-based modeling, in 12th IEEE-RAS International Conference on Humanoid Robots (Humanoids), pp , [2] J. Maitin-Shepard, M. Cusumano-Towner, J. Lei, and P. Abbeel, Cloth grasp point detection based on multipleview geometric cues with application to robotic towel folding, in IEEE International Conference on Robotics and Automation (ICRA), pp , May [3] P. C. Wang, S. Miller, M. Fritz, T. Darrell, and P. Abbeel, Perception for the manipulation of socks, in IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp , [4] Y. Li, Y. Yue, D. Xu, E. Grinspun, and P. K. Allen, Folding deformable objects using predictive simulation and trajectory optimization, in IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp , IEEE, [5] J. Stria, D. Prusa, V. Hlavac, L. Wagner, V. Petrik, P. Krsek, and V. Smutny, Garment perception and its folding using a dual-arm robot, in IEEE International Conference on Intelligent Robots and Systems (IROS), pp , IEEE, [6] G. T. Zoumponos and N. A. Aspragathos, A fuzzy strategy for the robotic folding of fabrics with machine vision feedback, Industrial Robot: An International Journal, vol. 37, pp , may [7] R. R. Ma, A. Spiers, and A. M. Dollar, M2 gripper: Extending the dexterity of a simple, underactuated gripper, in IEEE International Conference on Reconfigurable Mechanisms and Robotics (ReMAR), [8] R. R. Ma, L. U. Odhner, and A. M. Dollar, A modular, opensource 3d printed underactuated hand, in IEEE International Conference on Robotics and Automation (ICRA), pp , IEEE, [9] Y. Tlegenov, K. Telegenov, and A. Shintemirov, An open-source 3d printed underactuated robotic gripper, in IEEE/ASME 10th International Conference on Mechatronic and Embedded Systems and Applications (MESA), pp. 1 6, IEEE, [10] A. G. Zisimatos, M. V. Liarokapis, C. I. Mavrogiannis, and K. J. Kyriakopoulos, Open-source, affordable, modular, light-weight, underactuated robot hands, in IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp , IEEE, [11] B. Ward-Cherrier, L. Cramphorn, and N. F. Lepora, Tactile Manipulation With a TacThumb Integrated on the Open-Hand M2 Gripper, IEEE Robotics and Automation Letters (RA-L), vol. 1, pp , Jan [12] M. Kaboli, K. Yao, and G. Cheng, Tactile-based Manipulation of Deformable Objects with Dynamic Center of Mass, in International Conference on Humanoid Robots (HUMANOIDS), pp , IEEE,