A System-Level Cooperative Multi-Agent GNSS Positioning Solution

    •  Greiff, M., Di Cairano, S., Kim, K.J., Berntorp, K., "A System-Level Cooperative Multi-Agent GNSS Positioning Solution", IEEE Transactions on Control Systems Technology, DOI: 10.1109/​TCST.2023.3307339, Vol. 32, No. 1, pp. 158-173, October 2023.
      BibTeX TR2023-135 PDF
      • @article{Greiff2023oct,
      • author = {Greiff, Marcus and Di Cairano, Stefano and Kim, Kyeong Jin and Berntorp, Karl},
      • title = {A System-Level Cooperative Multi-Agent GNSS Positioning Solution},
      • journal = {IEEE Transactions on Control Systems Technology},
      • year = 2023,
      • volume = 32,
      • number = 1,
      • pages = {158--173},
      • month = oct,
      • doi = {10.1109/TCST.2023.3307339},
      • url = {}
      • }
  • MERL Contacts:
  • Research Areas:

    Control, Dynamical Systems, Signal Processing


We present a multi-agent cooperative estimation method for improving the performance of global navigation satellite systems (GNSSs). The proposed method uses existing receiver technology, avoids inter-agent communication, and min- imizes the computational overhead in the agents. The method is based on recursive mixed-integer Kalman filtering for a system characterized by several agents in a bipartite star graph structure, where the nodes in one of the vertex sets perform local filtering based on local information, and a single node in the other vertex set estimates all of the system states using inter- agent error correlations in the context of partially overlapping local state spaces. We conduct extensive Monte-Carlo simulation studies in an urban driving scenario using a road map from an actual city, incorporating real satellite trajectories and realistic ionospheric bias modeling. In addition, we perform a hardware- in-the-loop study. The results indicate that the method can correct erroneous estimates in faulty agents by leveraging cooperation with other agents, improving accuracy from decimeter level to centimeter level for that particular agent. When all agents have similar residual biases, expected improvements in the root-mean- square position error typically range between 20–100%.