Robots that localize radio sources and navigate with wireless signals, from probabilistic RF localization to a 10-GHz robotic testbed.
Overview
Wireless signals carry rich spatial information in their multipath, and a robot can use them to find a radio source and navigate toward it, even when cameras and maps are incomplete. But a single RF measurement is usually ambiguous: blockage and reflections leave several plausible source locations.
We treat this as probabilistic inference: learn full belief maps over where a transmitter could be, use those beliefs to guide a robot’s search, and validate everything on a 10-GHz robotic testbed in the upper mid-band (FR3).
Rather than forcing an ambiguous RF snapshot into one coordinate, we learn full posterior distributions over transmitter location, preserving directional and multimodal uncertainty. Our methods match measured multipath against ray-traced wireless digital twins, learn measurement-conditioned likelihoods, jointly infer position and heading, and keep working when the robot has mapped only part of its environment.
Belief maps become guidance for a robot's search policy. Using ray-traced digital twins as a physics-informed prior, our reinforcement-learning agents navigate unseen indoor environments toward a transmitter zero-shot, with the robot's SLAM system supplying the map and pose.
MobiFR3 is our 10-GHz robotic RF testbed. It pairs Pi-Radio SDR channel sounding and Vivaldi antennas with TurtleBot4 / ROS 2 robots, LiDAR, odometry, and RGB sensing, so every RF measurement is tied to a precise robot pose. Measurement-aligned Sionna RT scenes calibrate simulation to the hardware, connecting simulation-trained models to physical experiments.
Figures and video courtesy of Haozhe Lei.
