Cognitive Hybrid Control of an Autonomous Agent

In this paper, we investigate the coupling of two different cognitive modules, within the framework of embodied cognition. On one hand, an artificial agent is let to interact with its environment in order to learn the prediction of multi-modal sensory situations. The agent, makes use of a forward model as a basic cognitive tool. The trained system learns to successfully predict a multi-modal sensory representation of obstacles, formed by visual and tactile stimuli. On the other hand we synthesize a Bayesian controller to produce an obstacle avoidance behaviour. Using a real robot, the trained forward model and the Bayesian controller are coupled to solve a low-level task. The forward model is fed a covert motor command. When the covert motor command produces a collision-free sensory prediction the motor command is executed. When this is not the case, the Bayesian controller produces a motor command providing a collision free sensory situation by means of probabilistic reasoning. Experiments show that the coupling allows the robot to safely navigate among obstacles in its environment.