Building Generalization Into Behavior Generation Via Adaptive Compositions of Regularities

2026-05-29

AI Summary

Existing robotic systems struggle to generalize effectively in dynamic environments due to a lack of prior knowledge about environmental structure. This work proposes a strong inductive bias based on adaptively composing environmental regularities through the AICON framework, which models predictable patterns as interaction processes within differentiable networks. By integrating sensory feedback with gradient-based optimization, the approach dynamically modulates the influence of each regularity to generate context-sensitive behaviors. Experiments demonstrate that the method consistently produces plausible actions across diverse unseen conditions, failing only when the underlying regularities themselves are absent. Ablation studies further confirm the adaptive selection of regularities, highlighting the framework’s inherent generalization capability.

Abstract

Generalization in robotics requires prior knowledge about how the world is structured, yet this structure changes from one situation to the next. This paper investigates the proposition that generalization arises from adaptively composing regularities -- predictable relationships within the robot-environment system -- into situation-appropriate structures for behavior generation. We examine this proposition by analyzing the mechanism in AICON (Active InterCONnect), a framework representing regularities as interacting processes in a differentiable network, where sensory feedback realizes composition and gradient descent generates behavior. To isolate adaptive composition as the key mechanism, we study a simple simulated problem in which all relevant regularities can be identified. We expose the resulting model to a wide range of novel conditions not considered during design, and we find that it generates context-appropriate behavior in all but one case, where encoded regularities are provably insufficient. Ablations reveal that the network automatically modulates which regularities influence behavior based on their informativeness. These results suggest that adaptive composition of regularities constitutes a powerful inductive bias for building generalization into behavior generation.

Problem

• generalization
• behavior generation
• regularities
• adaptive composition
• robotics

Innovation

• adaptive composition
• regularities
• generalization
• differentiable network
• behavior generation