Distributed optimization methods for environment discovery by autonomous agents – ODISSEE

Consider a distributed system formed by autonomous mobile agents seeking to simultaneously discover their physical environment and planify their motions in a coordinate fashion. Scenarios of interest include autonomous underwater vehicles (AUVs) and drones.

Unlike the standard scenario which assumes that some fusion center is able to collect the sensors' observations and to govern the global network's behavior, we assume on the opposite that the processing and the decisions are achieved at the agent level. To accomplish the global network's mission, individual nodes are able to exchange limited information with their neighbors through unperfect wireless links and in an asynchronous fashion.

The ODISSEE problems addresses two crucial problems regarding defence applications: i) Distributed estimation and communication protocols in wireless sensor networks, ii) Planification, optimization and coordination in distributed mobile systems. The rationale behind ODISSEE relies on the claim that the above objectives share a common formulation in terms of a distributed optimization problem. The aim of ODISSEE is to bridge the gap between theoretical optimization methods and practical defence applications related to statistical estimation and motion coordination.

In a first stage, our methodological approach shall consist in the construction and the analysis of new distributed optimization algorithms. The ambition is to extend powerful optimization methods to a distributed framework involving unperfect communications and asynchronous protocols.

In a second stage, our findings will be applied to motion coordination and statistical environment estimation. As far as motion coordination is concerned, our distributed algorithms shall allow the agents to optimize their trajectories while keeping specified navigation constraints satisfied, such as rendez-vous, flocking, coverage optimization or global connectivity maintenance. As far as statistical estimation is concerned, it is generally hopeless to expect that the environment can be represented by a parametric model on which standard maximum likelihood methods could be applied. More evolved kernel-based machine learning techniques shall be applied.

Practical efficiency of our methods will be validated by numerical experiments. We shall consider practical scenarios where agents must simultaneously learn their environment and navigate without any prior knowledge on the latter.