next up previous contents
Next: 6.2.1 Robot setup Up: 6. Visuomotor model for Previous: 6.1.2 High-dimensional image data

6.2 Methods

First, the robot collected samples of images of an object and corresponding grasping postures. Afterward, training patterns were obtained by processing the images and coding the posture angles redundantly using tuning curves. The resulting distribution of patterns was approximated by a mixture of local PCA. After presenting an object, based on the approximation, a grasping posture could be recalled.


Heiko Hoffmann