Biologically-inspired 3D grasp synthesis based on visual exploration

Abstract

Object grasping is a typical human ability which is widely studied from both a biological and an engineering point of view. This paper presents an approach to grasp synthesis inspired by the human neurophysiology of actionoriented vision. Our grasp synthesis method is built upon an architecture which, taking into account the differences between robotic and biological systems, proposes an adaptation of brain models to the peculiarities of robotic setups. The architecture modularity allows for scalability and integration of complex robotic tasks. The grasp synthesis is designed as integrated with the extraction of a 3D object description, so that the object visual analysis is actively driven by the needs of the grasp synthesis: visual reconstruction is performed incrementally and selectively on the regions of the object that are considered more interesting for grasping.

Towards Grasp-Oriented Visual Perception for Humanoid Robots

“…To solve these problems, recently, Racatalá et al 23 developed a model of vision-based grasping (VBG) following a sense-planact paradigm. The system is based on the early separation and late integration of visual analysis through the two visual streams.…”

“…The GOVP system, as shown in Fig. 1, is constructed using a filter based architecture (FBA) initially proposed by Recatalá et al 23 It consists of three types of basic components: (i) hardware components (bullet-shape) including sensors and actuators, (ii) virtual filters (rectangular) that handle operations such as feature extraction or a control law, and (iii) data sets (ellipsoid) that store groups of data produced and processed by the above modules. A task is realized through a set of connected model components that are simultaneously active, where the data sets constitute an internal, non-centralized memory spreading along the chain of processors.…”

“…23 Along the dorsal stream, grasp-relevant visual features of an object are extracted and processed (CIP) Towards Grasp-Oriented Visual Perception for Humanoid Robots 395 in order to select and evaluate (AIP) a set of appropriate grasp configurations (or affordances). At the same time, the object identification process that is performed along the ventral pathway can further bias the grasp selection through its memory of previous experiences.…”

A strategy for grasping unknown objects based on co-planarity and colour information

“…The body of work in the area of robotic grasping is significant, see, e.g., [1][2][3][4][5][6][7][8][9][10]. We distinguish approaches based on the level of a-priori object information used to model the grasping process.…”

“…We distinguish approaches based on the level of a-priori object information used to model the grasping process. In particular, objects to be grasped may be assumed known, that is, both the shape and the appearance of the object are known and used to associate specific grasping strategies to them through exploration, (see, e.g., [2,3]) or different types of supervised learning (see, e.g., [9,10]). When objects are assumed to be unknown (as in our case), the assumptions of the system naturally need to be much more general in order to generate suitable grasping hypotheses (see, e.g., [4]).…”