Inference of affordances and active motor control in simulated agents

• DOI: 10.3389/fnbot.2022.881673
• Publication Date: 2022-08-11T12:51:11Z
• Authors: Fedor Scholz; Christian Gumbsch; Sebastian Otte; Martin V. Butz
• Keywords: Affordance

Flexible, goal-directed behavior is a fundamental aspect of human life. Based on the free energy minimization principle, theory active inference formalizes generation such from computational neuroscience perspective. theory, we introduce an output-probabilistic, temporally predictive, modular artificial neural network architecture, which processes sensorimotor information, infers behavior-relevant aspects its world, and invokes highly flexible, behavior. We show that our trained end-to-end to minimize approximation energy, develops latent states can be interpreted as affordance maps. That is, emerging signal actions lead effects dependent local context. In combination with inference, invoked, incorporating As result, simulated agent flexibly steers through continuous spaces, avoids collisions obstacles, prefers pathways goal high certainty. Additionally, learned suitable for zero-shot generalization across environments: After training in handful fixed environments obstacles other terrains affecting behavior, it performs similarly well procedurally generated containing different amounts various sizes at locations.