In-hand Manipulation for Active Object Recognition

Visual object recognition systems mounted on the automatic mobile agents face the challenge of unconstrained data, but simultaneously can realize the object recognition task in the human robot interactive environment by actively exploring. In this paper, we propose an active object recognition method to actively explore the more interesting positions of observation to reduce the uncertainty of object recognition. We propose to use the bayesian formula to iteratively accumulate the experience of image perception, and actively explore better recognition positions through control strategies. It is similar to humans actively exploring in the cognitive process, establishing the coupling relationship between perception and action, which is conducive to selecting the suitable posture to predict the labels. Furthermore, the effects of inhibiting action exploration behavior are analyzed with passive strategies. The results from the recognition of GERMS confirm that the proposed method successfully learns the effective category identification strategy, and the active selection action further boosts the recognition performance.