With the current achievement on multirobot localization algorithm, our next step is to generalize the current algorithm to enable the multirobot system in reality.
The work has been accepted by ISRR [1] and submitted to ICRA [2].
In the proof of the boundedness of DKF, the assumption that all the local KF converges is critical. While the measurement exchange rather than the diffusion step that really contributes to the estimation accuracy, we are interested in whether the message exchange can be further minimized, especially in energy critical scenarios.
We can begin with a simple case. Consider a time-varying scalar x, and several nodes are trying to estimate (or measure for simplicity) x and each nodes has different accuray. If all nodes can observe x, the DKF algorithm converges unquestionably. However, what is the best diffusion scheme that makes the overall covariance minimized. In other case, when some nodes can not observe x, the convergence requires that such nodes should receive the measurement from other nodes. However, while those nodes seem like a information user rather than a provider, it is beneficial to the overall covariance performance if other nodes use their estimate.
We begin with several basic topologies before arriving ultimate general conclusions.
Star topology
The next step is to extend the above result on more general case where x itself is a vector, and each nodes
Topology is one of the main factor in multirobot or multiagent system. While the robots that we consider are mobile most of the time, the underlying topology varies with time correspondingly. The effect of varying topology is therefore of great importance.
In [1], the varying topology is analyzed. While this work provides a framework to deal with varying topology, the model in [1] is only consensus type data fusion. Whether is can be extended to other types of data fusion, e.g. Kalman filter as our main focus, is the goal of this investigation.