Page 164 - Human Inspired Dexterity in Robotic Manipulation

Dynamic Manipulation Based on Thumb Opposability 161

shifted the tip of the phantom to generate the translational motion, as shown
in Fig. 8.11B, where the white box surrounded by the solid line and gray
box surrounded by the dashed line are the positions of the object in the cur-
rent and previous frames, respectively. In both cases, the hand followed the
operator’s command, and stably manipulated the objects.
Besides, a procedure of autonomously picking up an object by the
dual-arm robot is shown in Fig. 8.12. To pick an object without human
intervention, (1) cognition is an intensely important part, and is different
from the former case of supervisory control. This step needs an object-pose
estimation and grasping-strategy engine, as shown in the bottom part of
Fig. 8.12. The object-pose estimation is used to recognize the approximate
center position and pose of the object randomly laid on a table. The grasping
strategy engine infers (1) grasp type such as pinch, n-finger, and envelop,
(2) grasp point, and (3) approach vector to the object from the extrinsic
information of the object such as pose, size, and shape. Actually, the grasping
strategy engine was implemented by learning methods using neural net-
works, and extension is also available. For these reasons such desired values,
grasp type, grasp point, and approach vector are autonomously inferred from
information measured by a 3D sensor without any human intervention.
After the step of (1) cognition, (2) reaching of the hand-arm robot to a target

Object pose estimation Grasping strategy engine Grasp/manipulation

Fig. 8.12 Autonomy in grasp and manipulation.