Table 4 IV: Classification Accuracy with Object Weight.
Related Figures (15)
Fig. 1: The sensorised kirigami gripper developed includes soft air pouches fixed to the contact surfaces for grasping force estimation. Load cells on the parallel jaw allow the actuation force and weight of the object to be measured, and a camera is mounted onto the base of the actuator system for keypoint detection. Junbang Liang, Joao Buzzatto, Bryan Busby, Ricardo V. Godoy, Saori Matsunaga, Rintaro Haraguchi, Toshisada Mariyama, Bruce A. MacDonald, and Minas Liarokapis TABLE I: Dimensions of the Kirigami Gripper. Fig. 3: The kirigami is marked with predicted keypoints, which are visible from the camera fixed to the base of the parallel jaw gripper. The predicted keypoints are numbered from | to 16, with points 1 to 6 corresponding to the outer vertices of the kirigami during grasping, and points 7 to 16 corresponding to the marked keypoints around the center cavity. Fig. 2: A fully-manufactured kirigami structure with air pouch sen- sors attached. The structure consists of a thin PET sheet sandwiched between two layers of Smooth-On Dragonskin-30 silicone rubber. The air pouches are manufactured with Dragonskin-10 silicone rubber, and are attached to the kirigami with adhesive. Refer to Table I for the dimensions indicated. TABLE II: Item Dimensions in the Food Object Set. Fig. 4: The experimental procedure for single-grasp object classifi- cation with the kirigami gripper. From left to right: (a) the gripper grasps the object on the platform; (b) the object is lifted upwards slightly, and several sensor readings are taken; (c) the gripper rotates by 90 degrees and takes weight measurement with the load cells. The kirigami conforms to different object sizes and ge- ometries, generating unique arrangements of the keypoints for each object’s shape and orientation allowing for object identification. Fig. 5: (a) The model objects used in the shape object set. The cylinders and rectangular prisms are made out of 3D-printed polylactic acid plastic and have diameters or widths of 20 mm, 40 mm, and 60 mm. (b) The food object set consists of bananas, silicone chicken nuggets, figs, large green grapes, small red grapes, ibuprofen pills, oranges, pears and tomatoes. (c) The transparent objects include two round bottles, a small cup, a square bottle, a test tube and a pipette. The square bottle and large round bottle are made of glass, while the remaining objects are made out of plastic. TABLE III: Item Dimensions in the Transparent Object Set. 1) Shape Object Set: The first object set consists of cylindrical and rectangular geometries, illustrated in Fig. 5a. These objects are fabricated using polylactic acid plastic through 3D-printing and vary in diameter (in the case of cylinders) or width (in the case of rectangular prisms). Further details on the dimensions of the objects are provided in Fig. 5a. These objects with uniform geometries provide a reliable control group for analysing the classification accu- racy of the proposed methodology. Unlike food items, which vary in surface geometry and dimensions, the structures of each shape model object differ only in terms of the overall diameter or width. Fig. 6: The experimental setup for measuring the force pressure response of the air pouch sensors. The testing rig consists of plates connected to an adjustable base with load cells, both of which have force resolutions of 0.25 g. The distance between the two plates is varied to be 20, 30, 40 and 50 mm apart to simulate different-sized objects. Fig. 7: Geometry variations of the plates used for measuring the force pressure response of the air pouch sensors with the flat surface shown in (a) and a curved surface shown in (b). measure the response of the air pouch sensors, we conducted grasping force experiments with the kirigami experimental setup, as illustrated in Fig. 6, pair of flat plates connected to an adjustable object sizes affect the force pressure response o gripper. The distance between the two plates gripper. The comprises a base via two oad cells. The base allows the distance between the two plates to be modified, enabling us to investigate how varying f the kirigami is varied to be 20, 30, 40 and 50 mm apart to simulate different-sized objects. Moreover, to study the impact of di shapes, we used two types of flat and curved c fferent object ontact plates. Fig. 7 displays the contact geometry of the plates. TABLE V: Classification Accuracy without Object Weight. TABLE VI: Classification Accuracy on the Food Dataset with Different Kirigami Material. Fig. 9: Plot showing the relationship between the forces exerted and the air pouch pressure measurements when performing grasp force experiments with different plate geometries. A clear distinction is observed between the grasping of the two surfaces, as the curved plate produces higher pressure readings than the flat plates. Fig. 8: Bar graph showing the importance of each measured parameter. The load cells’ weight and force readings are of the highest importance when classifying objects. The pressure readings of the air pouches, the y-positions of keypoints 2 and 5 and the x-positions of keypoints 7-15 are of secondary significance. Fig. 10: Scatter plot showing force exertion and air pouch pressure measurements from the grasp force experiments with varying plate separation distances. The plot indicates that the force pressure responses remain consistent across different plate separations.
