RF-Based 3D SLAM Rivaling Vision Approaches

We present CartoRadar, a novel RF-based SLAM system that delivers high-fidelity 3D mapping with centimeter-level accuracy. CartoRadar builds on top of the advancements in learning-based RF imaging. However, learning-based systems often exhibit variation in prediction accuracy during inference. To address this challenge and enable robust RF sensing, CartoRadar introduces a novel, training-free uncertainty quantification method tailored to RF signals. Additionally, CartoRadar features an efficient SLAM algorithm that incorporates this uncertainty into the mapping process. We deploy CartoRadar on a mobile robot and conduct extensive evaluations across 14 floors in 5 buildings. Results show that CartoRadar achieves a trajectory error of 14.1 cm, outperforming camera-based baselines by 72.1%. For mapping, CartoRadar achieves an accuracy of 7.4 cm and a completion of 8.1 cm, improving over vision methods by 46.2% and 67.6%, respectively.

This figure illustrates the system architecture of CartoRadar, consisting of two components:

1. Robust RF Sensing with Uncertainty Quantification takes in the raw RF signals and outputs range images with uncertainty estimates for each pixel.
2. Uncertainty-aware RF-Based SLAM takes in the predicted range images and incorporates the uncertainty estimates into the SLAM process. We propose a novel implicit occupancy field with probabilistic learning, effectively improving the accuracy, robustness and efficiency.

We compare our proposed uncertainty quantification methods with three baselines and report the NLL and AUSE results in the table above. Our method can be applied out of the box, without any re-training, and still achieves better performance than the traditional baseline methods. Furthermore, with one additional training step, our method can achieve even greater performance improvements.

We present the mapping results (left) and detailed views (right) of CartoRadar across different environments. For each trajectory, our results are compared with the corresponding LiDAR mesh map. The maps are color-coded according to surface normals, and the trajectories are shown as yellow lines. We also provide accompanying videos for better visualization and zoomed-in views of details.

Our system is evaluated over 14 floors across 5 separate buildings. The diversity of the buildings is notable, featuring distinct designs and materials with construction dates stretching from 1906 to 2006. The entire dataset was collected traversing a distance of 1527 meters. After signal processing, it consists of 6637 synchronized RF and LiDAR frames, respectively, aggregating to 223 GB.

We release code and dataset to facilitate future research in this direction. Detailed instructions can be found in the README of the code repo.