Federated Data-Driven Kalman Filtering for State Estimation

TL;DR

The paper tackles accurate ego-vehicle localization under privacy and communication constraints by introducing FedKalmanNet, a federated learning framework that extends KalmanNet with adapt-then-combine training. Each vehicle locally trains KalmanNet to estimate KF uncertainty matrices using private data, and a central server aggregates only model weights to form a global KalmanNet, preserving data privacy. Empirical results in the CARLA simulator show FedKalmanNet achieving near-centralized performance and outperforming collaborative decision-making baselines that fuse raw measurements, while requiring substantially less V2X communication. The approach enables scalable, data-driven localization with interpretable Kalman-based updates and offers a practical privacy-preserving alternative for autonomous driving deployments.

Abstract

This paper proposes a novel localization framework based on collaborative training or federated learning paradigm, for highly accurate localization of autonomous vehicles. More specifically, we build on the standard approach of KalmanNet, a recurrent neural network aiming to estimate the underlying system uncertainty of traditional Extended Kalman Filtering, and reformulate it by the adapt-then-combine concept to FedKalmanNet. The latter is trained in a distributed manner by a group of vehicles (or clients), with local training datasets consisting of vehicular location and velocity measurements, through a global server aggregation operation. The FedKalmanNet is then used by each vehicle to localize itself, by estimating the associated system uncertainty matrices (i.e, Kalman gain). Our aim is to actually demonstrate the benefits of collaborative training for state estimation in autonomous driving, over collaborative decision-making which requires rich V2X communication resources for measurement exchange and sensor fusion under real-time constraints. An extensive experimental and evaluation study conducted in CARLA autonomous driving simulator highlights the superior performance of FedKalmanNet over state-of-the-art collaborative decision-making approaches, in localizing vehicles without the need of real-time V2X communication.

Figures (5)

• Figure 1: The proposed end-to-end FedKalmanNet approach. The methodology consists of two key steps: adaptation and combination. During the adaptation step, each agent trains the corresponding local KalmanNet model (\ref{['first_eq_knet_FL']}) based on KF concept. This ensures that the local model will capture the unique characteristics of each agent's dataset. During the combination step, the central server aggregates the local KalmanNet models using a fusion rule (\ref{['combined_FL']}), thereby creating a global model that encapsulates the knowledge from all participating agents.
• Figure 2: Clients' trajectories from TownMap10 of CARLA
• Figure 3: Convergence of FedKalmanNet to CentrKalmanNet after 20 communication rounds
• Figure 4: Cumulative distribution function of ego vehicle localization accuracy
• Figure 5: FedKalmanNet outperforms the baseline methods, exploiting only self GNSS and velocity. LKF-SA has to integrate greater amount of information from neighbors to reach FedKalmanNet's accuracy