milliMap: Robust Indoor Mapping with Low-cost mmWave Radar

Abstract

Single-chip Millimetre wave (mmWave) radar is emerging as an affordable, low-power range sensor in automotive and mobile applications. It can operate well in low visibility conditions, such as in the presence of smoke and debris, fitting the payloads of resource-constrained robotic platforms. Due to the nature of the sensor, however, distance measurements are very sparse and affected by multi-path reflections and scattering. Indoor grid mapping with mmWave radars has not been yet explored. To this extent we propose milliMap, a self-supervised architecture for creating dense occupancy grid maps of indoor environments from sparse, noisy mmWave measurements. To deal with the ill-constrained sparse-to-dense reconstruction problem, we leverage the Manhattan world structure typical of indoor environments to introduce an auxiliary loss that encourages generation of straight lines. With experiments in different indoor environments and under different conditions, we show the ability of milliMap to generalise to previously unseen environments. We also show how the reconstructed grid maps can be used in subsequent navigation tasks.