Communication rates impact cooperative multi-agent tracking accuracy
Measuring the Value of Communication in Multi-Agent Tracking Under Finite-Rate Belief Sharing
Information Theory
Summary
Tracking a moving target using multiple agents working together requires them to share information. This paper shows that sharing too little information makes tracking worse, but sharing a moderate amount greatly improves it. Sharing even more data beyond that helps only a little, meaning there is a sweet spot for communication rates. The authors explain how communication limits and sensor arrangement together affect tracking performance.
What this means in practice
- •For autonomous vehicle teams: Optimize inter-vehicle communication bandwidth to improve collective tracking of dynamic objects with minimal needed data rates.
- •For robotics system integrators: Configure communication protocols to balance tracking accuracy and data transfer costs in sensor networks using extended Kalman filtering fusion.
Authors
Danial Dehghani, Maximilian Lübke, Zahra Seifaei, Norman Franchi
Abstract
This paper investigates the value of communica- tion (VoC) in cooperative multi-agent tracking under rate- constrained information exchange. Although prior work has studied distributed tracking, quantized communication, and robust fusion, it remains unclear at what communication rate most of the achievable tracking gain is obtained, beyond which additional communication resources yield only marginal benefit. We consider stationary agents that observe a moving target through heterogeneous range and bearing measurements, perform local extended Kalman filtering (EKF), quantize their posterior state means, and fuse exchanged information via covariance intersection (CI). Performance is evaluated versus communication rate relative to no-communication baselines, a centralized oracle, and a CI-perfect reference. The results show that communication is not uniformly beneficial: very low rates can degrade performance due to coarse quantization, whereas moderate rates recover most of the achievable gain. In the considered scenario, increasing the rate from 4 to 12 bit/s per agent reduces the root-mean-square error (RMSE) from 13.93 m to 6.35 m, while a further increase to 32 bit/s per agent improves it only marginally to 5.93 m. This reveals a clear knee region at moderate rates, beyond which the remaining loss is mainly due to the conservatism of the fusion architecture. Overall, the proposed framework characterizes low-rate degradation, communication-efficient operation, and diminishing returns, while highlighting the joint role of rate and sensing geometry in determining communication benefit.