Date of Original Version
© 2011 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works.
Abstract or Description
We investigate mobile ad-hoc indoor networks consisting of simple inexpensive robots, LANdroids, with limited wireless communication range and without any range or location sensors. We focus on the problem of using the mobile LANdroids to take responsibility for maintaining connectivity between a static Gateway and mobile Targets that move beyond the communication range of an established network. We refer to such a tracking task as Target Tethering. This type of network commonly uses IEEE 802.11 wireless protocols for communication, with Received Signal Strength Indicator (RSSI) as a measure of radio signal strength. RSSI data is noisy and poorly relates to distance in indoor environments, leading to a challenging Target Tethering task. Some algorithms use the trace of single-source RSSI data to infer distance between two nodes and use it to compute a Target Tethering policy. However, such distance estimates are poor. We instead aim at inferring physical network layout from RSSI data among multiple nodes. We introduce a novel approach based on Cluster Geometries, classes of network nodes corresponding to rotation-invariant physical layouts of LANdroids and a mobile Target, with the conjecture that multi-robot RSSI data can distinguish the Cluster Geometries and therefore the physical layouts. We proceed with extensive experiments and support our conjecture by showing successful classification of the designed Cluster Geometries given the multi-robot RSSI-based data. We then combine the estimated Geometries with motion patterns of the moving Targets to show that suitable multi-robot Target Tethering policies for unknown indoor environments can be learned using multi-agent reinforcement-learning. Specifically, we use an interesting variation of Q-learning where we first learn offline base policies in general open environments and later specialize the policies seamlessly during online execution to account for obstacles in the indoor environment.
Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2011, 2327-2332.