Shape constancy in action and perception

点击翻译按钮获取中文摘要

The ability to make accurate, dexterous hand movements for grasping and manipulating objects is a defining behavioural feature of humans. These actions typically involve little mental effort, giving rise to the `common-sense' impression that the interaction between the visual and motor systems that underlie them is quite straightforward: i.e., we see the target object, and we grasp it. But things are not that simple. Consider, for example, the task of reaching for our smart phone. We would be at a loss if asked to specify the geometry that gets the fingers where they need to be. At minimum, we would need a description of the phone's structure and the distances and angles from the phone's surface to our limb and its various segments. And yet, when we reach for our phone in a myriad of everyday different contexts, our visuomotor system samples the visual array and performs the necessary computations to initiate the reach in a fraction of a second. Reaching for an object is a complex process that operates at the intersection of a host of systems involving vision, haptics, attention, memory, motor, and executive control. The purpose of this research program is to further explore how well geometric descriptions of the structure of 2D and 3D objects can explain how we perceive and act on objects across a range of viewpoints and viewing distances. In the short term, I propose: (1) to examine how well grasp-point selection is explained and guided by geometric descriptions of 2D and 3D object shape, such as the medial axis and its associated radius/distance function; (2) to develop and test a model of grasp point selection that is based on these geometric descriptions; and (3) to use MRI-guided transcranial magnetic stimulation to see what role the lateral-occipital and anterior intraparietal cortices play in computing the medial axis for forced-choice object classification tasks and the selection of stable contact points for grasping. This research program adopts a model of the visual system that uses the medial axis transform as a basis to derive structural representations and properties that assist with identifying and recognizing objects and with the selection of grasp points. Due to its computational nature and its resonance with computer vision and image processing, a model of grasp point selection based on the medial axis stands as a promising avenue for application in robotics and artificial intelligence, in addition to providing new directions for basic research using EEG, neuroimaging, and single-unit recordings. The long-term goals of this research program are geared towards the design and implementation of robots capable of flexible, human-like grasps and manipulation that can be deployed in environments ranging from assisted care to those that are hazardous and/or inaccessible to humans and, in particular, where the manual manipulation of novel objects of unpredictable shape and size is required.