Reproduce “Inference of Affordances and Active Motor Control in Simulated Agents” on OpenResearch

About this paper

Fedor Scholz, Christian Gumbsch, Sebastian Otte, Martin V. Butz

Researchers at the University of Tübingen developed an active inference framework that integrates learned affordance maps to enable computationally efficient, flexible, and robust goal-directed behavior in continuous environments. The approach demonstrated zero-shot generalization to novel environments and effective uncertainty avoidance capabilities.

Key insights

• Integrating learned affordance maps into active inference can overcome computational intractability by constraining the planning search space in complex environments.
• Behaviorally relevant representations (affordance codes) can emerge through end-to-end training that optimizes for accurate action-effect prediction.
• The proposed framework enables robust goal-directed control and strong zero-shot generalization capabilities in continuous, dynamic environments.

Summary by alphaXiv.