Multi-Agent Navigation with Reinforcement Learning Enhanced Information Seeking

Multi-agent robotic networks allow simultaneous observations at different positions while avoiding a single point of failure, which is essential for emergency and time-critical applications. Autonomous navigation is vital to the task accomplishment of a multi-agent network in challenging global navigation satellite systems (GNSS)-denied environments. In these environments, agents can rely on inter-agent measurements for self-positioning. In addition, agents can conduct information seeking, i.e., they can proactively adapt their formation to enrich themselves with position information. Classical signal processing tools can efficiently exploit the knowledge of system and measurement models, but are not applicable for long-term objectives. On the other hand, data-driven approaches like reinforcement learning (RL) are suitable for long-term action planning but have to face the critical curse of dimensionality. In this paper, we propose a multi-agent navigation scheme with RL-enhanced information seeking, which simultaneously takes advantage of model-based and data-driven approaches to collaboratively accomplish challenging objectives while exploring a GNSS-denied environment.

I. Introduction

Multi-agent networks have attracted an ever increasing attention in sensing and exploration applications, thanks to the increased exploration efficiency due to collaboration and the capability of observation from different points of view. The ability of ubiquitous navigation is essential for a multi-agent network. A typical navigation problem involves positioning, i.e., estimating the position of an entity like a vehicle, human or robot, and goal approaching, i.e., guiding this entity from one place to another. As specific to multi-agent navigation, collaboration among agents enhances the navigation capability by cooperative positioning and formation optimization [1], [2]. Eventually, for multi-agent information seeking [3], [4], agents can proactively adapt their formation, so that the positioning uncertainty is actively minimized while approaching a goal.