Sensorimotor Coordination in a Humanoid Robot: Building Intelligence on the iCub - Bioinspired Approaches for Human-Centric Technologies - page 189

Fig. 6.15 Bottom left : the learned map between object curvatures and successful grasps. Curva-
ture is normalized between 0 and 1. Points with curvatures in the range of 0.1 and 0.2 are
preferable as they are likely to bring about successful grasps
6.6.4.3 Complete Pipeline
To evaluate the complete system, we define a performance indicator as follows: a
grasp is successful if the robot can hold it firmly without dropping it. The exper-
iment we present here also demonstrates that the curvature map (learned earlier)
generalizes to novel objects. We execute 20 trials on each object in the test set
(Fig. 6.16 ), i.e., 80 trials in total, achieving an overall success rate of 91.25 %. Such
accuracy makes these algorithms suitable to be employed in robust manipulation
tasks. As grasping lacks a standardized benchmark, we compare our approach with
a simple top grasp (grasping the object always from the top), which has been widely
used in the past. We tested this more stereotyped grasp on the same objects showed
in Fig. 6.16 carrying out 20 trials per object as before. The success rate for the
cylinder and the bottle was significantly lower (15 % for both) quite obviously
because of their elongated shape that makes the top grasp unsuitable. We also
achieved 65 % for the stuffed dog and 80 % for the cube. These results confirm the
considerable performance gain of the complete grasping pipeline.
6.7 Vision
The remaining element in this journey through the structure of the iCub controllers
is certainly vision. We strive to provide reliable estimates of the position and shape
of objects in space since this enables the control of action as presented earlier