Neurally Augmented State Space Model for Simultaneous Communication and Tracking with Low Complexity Receivers

In this paper, we propose an integrated sensing and communications (ISAC) system where a base station (BS) equipped with an antenna array and a co-located radar receiver transmits data packets while simultaneously tracking the position of users. We restrict our attention to the simplest hardware architecture, where the beamforming array can generate beams from a discrete codebook and the receiver is equipped with a single analog to digital converter, thereby allowing for scalar-only measurements where angular information is lost. Under such restrictive constraints, the observation likelihoods are hard to model, which motivates us to learn them via neural networks. This learned likelihoods are then incorporated into a state space model where Bayesian filtering can be performed. We test our method in complicated road geometries and show that our tracker is capable of following high mobility users most of the time. Furthermore, when the track of a user is lost, it often takes only a few measurements until is is recovered, disposing of the need for time consuming beam alignment procedures.