Grasping with flexible viewing-direction with a learned coordinate transformation network

We present a neurally implemented control system where a robot grasps an object while being guided by the visually perceived position of the object. The system consists of three parts operating in a series: (i) A simplified visual system with a what-where pathway localizes the target object in the visual field. (ii) A coordinate transformation network considers the visually perceived object position and the camera pan-tilt angle to compute the target position in a body-centered frame of reference, as needed for motor action. (iii) This body-centered position is then used by a reinforcement-trained network which docks the robot at a table so that it can grasp the object. The novel coordinate transformation network which we describe in detail here allows for a complicated body geometry in which an agent´s sensors such as a camera can be moved with respect to the body, just like the human head and eyes can. The network is trained, allowing a wide range of transformations that need not be implemented by geometrical calculations