Kolmogorov–Arnold Controllers: Toward Smooth and Safer Embedded Policies

Swarnava Dey

Keywords: Kolmogorov–Arnold Networks, Embedded Reinforcement Learning, Safe Control, Policy Learning, Neural Controllers

TL;DR: We show that Kolmogorov–Arnold Networks can act as smooth, compact policy controllers that improve safety behavior on embedded RL tasks without increasing model footprint.

Abstract: Reinforcement learning (RL) controllers often rely on multilayer perceptrons (MLPs), whose sharp action responses can behave unpredictably when safety intervention is required. Kolmogorov–Arnold Networks (KANs), built from smooth spline operators, offer a more regular alternative. We compare MLP and KAN policies under two RL algorithms and two control tasks of differing difficulty, using identical training setups and a light supervisory mechanism that overrides actions near unsafe conditions. In the easier task, both architectures achieve comparable performance. In the harder, safety-critical setting, the MLP controller becomes unstable under supervision, while the KAN policy maintains consistent learning and achieves substantially lower unsafe-state rates. Our goal is not to claim superiority of either model, but to characterize their stability, safety profiles, and design trade-offs in the context of embedded AI controllers. Early results suggest that mid-sized KANs produce smoother activation patterns and safer trajectories while retaining sample efficiency comparable to equally parameterized MLPs.

Submission Number: 12