Learning Where to Park

We consider active inference as a novel approach to the design of synthetic autonomous agents. In order to assess active inference’s feasibility for real-world applications, we developed an agent that controls a ground-based robot. The agent contains a generative dynamic model for the robot’s position and for performance appraisals by an observer of the robot. Our experiments show that the agent is capable of learning the target parking position from the observer’s feedback and robustly steer the robot toward the learned target position.