CA-NKCF: Neural Kalman filter beats traditional estimators without covariance knowledge

In a new paper submitted to an IEEE journal, Stamatelis et al. propose the Covariance-Agnostic Neural Kalman Consensus Filter (CA-NKCF), addressing the challenge of online latent state estimation in distributed multi-agent systems. The filter operates without any knowledge of process or measurement noise statistics—hence "covariance-agnostic"—by combining prior estimates, learned consensus weights, and Kalman-like recursive updates. This hybrid architecture marries partial domain knowledge (e.g., motion models) with the representation power of deep neural networks, enabling agents to collaborate and exchange information for decentralized inference.

Extensive experiments on linear dynamics, the chaotic Lorenz system, and practical wireless tracking environments demonstrate CA-NKCF’s superiority. It consistently outperforms traditional distributed Kalman filters (DKF), particle filters, and purely model-free deep neural networks, even when the underlying motion or observation models are misspecified. The performance advantage remains stable across varying noise levels, random communication topologies, latent state dimensions, and observation clutter induced by scattering objects. With 13 pages and 9 figures, the work is under review at an IEEE journal and represents a significant step toward robust, scalable distributed estimation for robotics, autonomous systems, and wireless sensing.