A self-prior in active inference lets an agent spontaneously reach toward a tactile stimulus
The authors propose a density model of an agent's own multimodal sensory experience, called the "self-prior," and embed it in an active inference framework based on the free energy principle, so that behavioral references arise purely from an intrinsic process minimizing mismatches between average past sensory experience and current observations; in a simulated environment the agent spontaneously reaches toward a tactile stimulus without external reward.