View-independent recognition of grasping actions with a cortex-inspired model

To recognize how people interact with objects is essential for humans and artificial systems like robots. However, this recognition task is difficult and requires the capturing of the details of effector and goal object under a wide range of image transformations, such as view or position changes. Here, we demonstrate how specific effector-object interactions can be efficiently recognized by a simple, biologically plausible neural model. In line with biological evidence, the model applies a view-based approach for the recognition of grasping sequences from videos. The model generalizes to untrained views by interpolation between stored example views. In addition, it presents a novel physiologically plausible mechanism to capture the spatial relationship between effector and object. The results support the view that where and how an object will be grasped by an agent can be predicted without estimation of the 3D structure of the scene.