Wireless Positioning and Channel Charting with NVIDIA Aerial Omniverse Digital Twin

Our previous work focused on using NVIDIA Aerial Omniverse Digital Twin channel impulse response (CIR) data for supervised channel state information (CSI)-based UE positioning. Supervised CSI-based UE positioning consists of training an NN with a mean squared error (MSE) loss that compares the predicted UE positions to ground-truth labels, i.e., true UE positions. In contrast to prior research focusing on supervised UE positioning, the integration of NVIDIA Aerial Omniverse Digital Twin ray-traced channels presents a unique advantage: the ability to simulate complex and large scenarios with UE movements, allowing us to rapidly gather large and realistic CSI datasets. Figure 1 shows the NVIDIA Aerial Omniverse Digital Twin simulation geometry and Figure 2 shows the resulting estimated UE position map.  

Our next steps will focus on using NVIDIA Aerial Omniverse Digital Twin complex frequency response (CFR) channels for channel charting (CC). CC is an emerging self-supervised method that learns a map and transforms it from high-dimensional CSI space to a low-dimensional representation called the channel chart. The key property of a channel chart is that UEs that are nearby will be nearby in the channel chart [1]. Although this technique excels at preserving local geometry, it distorts global geometry, resulting in UE positions in arbitrary coordinates. To address this constraint, we have proposed CC in real-world coordinates [2]. This work leverages the fact that BST positions are known after their deployment and that strong received power at a certain BST coming from a certain UE indicates that the UE is in close vicinity to the BST. This technique enables one to perform approximate positioning but in a fully self-supervised manner. In addition, we have proposed a novel streaming CC architecture [3] that deals with the issue of storage limitations, as high-dimensional CSI is generated at extremely fast rates. This architecture maintains only a small set of representative CSI data, from which channel charts are learned.

Please visit the website for more details on another recent collaboration with NVIDIA, where we deployed a 5G virtual radio access network (vRAN) system built upon external pageNVIDIA Aerial Research Cloudcall_made platform and the external pageCC resources pagecall_made for more details on recent research in CC.