MIRROR: Mirroring Teammate Knowledge for Multi-Robot Exploration under Communication Constraints

Abstract

Teams of robots explore an unknown environment faster than a single robot only if they tell each other what they have seen. Most exploration systems assume that two robots in contact merge their maps completely, and on real platforms that assumption is not always feasible, since a link often carries far less than a full map. This thesis therefore takes a communication budget as its central assumption. A contact then delivers a small part of what a robot knows, so transmitting becomes selecting, and in deployed systems that selection falls to the order in which the map sits in memory. The thesis formalizes this problem and reduces it to one choice, the order in which the unsent cells travel. It then presents MIRROR, a policy that makes that choice deliberately. The sender maintains an estimate of what the receiver already knows, built from the cells it has delivered and the positions at which it has observed that teammate, and it sends first the cells beyond the edge of that estimate. Existing answers to a thin channel shrink the payload or negotiate its content over the channel itself, so the choice is either left undesigned or paid for in the bandwidth it was meant to save. MIRROR needs no extra message and no reply, so the estimate costs the channel nothing, and it fits any system that already exchanges positions. MIRROR is evaluated in simulation on six real building floor plans, one of them held out until a single blind run, with teams of two to seven robots, sensor ranges from 5 to 40 m, and budgets spanning two orders of magnitude, against policies from silence to unlimited exchange and an ordering that reads the receiver’s true map. Aiming the messages makes a team explore faster and waste less. With five percent of a full exchange per contact, MIRROR recovers between 93 and 98 percent of what communication can give on every floor tested, returns between a quarter and a half of the time a silent team spends, and removes nearly all of the duplicated sensing that consumes between 44 and 64 percent of a silent team’s effort.