Identifying and Localizing Dynamic Affordances to Improve Interactions with Other Agents in Continuous Environments

Allowing autonomous agents to learn by themselves to interact with other agents requires that they be able to recognize each other and be capable of inferring their behaviors. To comply with artificial developmental learning, we follow the radical interactionism hypothesis, in which an agent has no a priori knowledge on its environment. A previous work has shown that the agent can learn to identify, localize, and predict movements of mobile elements, but was only tested in discrete environments, limiting their applicability on real-world systems. This paper presents new mechanisms for the identification and localization of mobile entities in a continuous environment. These mechanisms learn the relations between the agent’s sensorimotor patterns and the entities, static or mobile, affording them, and store discovered properties in data structures called Signatures. The properties of signatures are then exploited to detect distant entities in surrounding environment without relying on a geometrical notion of space. These mechanisms were tested in a simple environment, the results showed how signatures integrate the complexity of a continuous environment through the limited sensory system of the agent, and localize distant entities through data structures that are compatible with previously developed behavior inference and prediction mechanisms.