Visual Checking of Grasping Positions of a Three-Fingered Robot Hand

Abstract

Abstract. We present a computer vision system for judgement on the success or failure of a grasping action carried out by a three-fingered robot hand. After an object has been grasped from a table, an image is captured by a hand camera that can see both the object and the fingertips. The difficulty in the evaluation is that not only identity and position of the objects have to be recognized but also a qualitative judgement on the stability of the grasp has to be made. This is achieved by defining sets of prototypic "grasping situations" individually for the objects.

Combining spatial and colour information for content based image retrieval

“…The local PCA-approach used in ColoSeek has been applied in previous work to classification of facial features [22], object-and texture recognition [23], hand posture recognition [24], and visual judging of grasp stability in robotics [25]. This variety shows the major benefit of the approach: it is inherently scale-and domain-independent.…”

“…Only algorithms which explicitly count the codeword access frequencies to avoid under-utilisation like ''Frequency Sensitive Competitive Learning'' [66,67] or ''Activity Equalisation VQ'' (AEV) [68] solve the problem without extensive parameter tuning. For ColoSeek, AEV is used because it was successfully applied to object recognition tasks earlier [23,25].…”

“…Due to the computational effort and the complexity of the domain, extensive search for suitable parameters is impossible, instead, algorithms have to work as ''black boxes.'' With the AEV-algorithm for clustering and SangerÕs method for PCA, as presented in the next two sections, two reliable tools with the desired properties are available, which have proven successful in several vision applications [22][23][24][25].…”

Acquisition and Application of a Tactile Database

“…slip detection or shape reconstruction), classification of tactile and haptic features as well as the feature extraction on tactile datasets and maybe even explorative strategies on the dataset. As a first application on this tactile database, an object (or rather object and posture) recognition with an universal classification architecture, which main appliance is the field of computer vision [23], [24], is tried. The approach is motivated in [25], for the current paper, we will give only a brief introduction.…”

“…By contrast, the third layer is trained supervised to the classification of these features. The main advantage of this architecture is its suitability for very high dimensional and complex data, as has been shown in several earlier applications both of the isolated stages [26] and as a whole [23], [24].…”