Dynamic learning of action patterns for object acquisition
Gabriele Peters, Thomas Leopold · International Journal of Intelligent Systems Technologies and Applications · 2007
We propose an active vision system for the acquisition of internal object representations. The core of the approach is an agent which learns goal-directed action patterns depending on the perceived environment via reinforcement learning. The user supervision is restricted to the definition of this goal in the form of a reward function. We demonstrate this approach by means of learning a strategy to scan an object. The agent moves a virtual camera around an object and is able to adapt her scan path dynamically to different conditions of the environment such as different objects and different goals of the data acquisition. The purpose of the acquisition which we consider here is the view-based reconstruction of non-acquired views. The scan pattern obtained after the learned path has stabilised allows a better reconstruction of unfamiliar views than random scan paths.