GeRaF reconstructs 3D geometry from radio signals, seeing through occlusion
New neural method achieves millimeter-level 3D reconstruction from RF signals.
GeRaF tackles a long-standing challenge: reconstructing precise 3D geometry from radio frequency (RF) signals. While RF sensing can penetrate occlusions (e.g., walls), its lensless nature introduces severe noise and specular reflections, making traditional volumetric rendering prohibitively expensive (cubic complexity). The team from MIT and EPFL (Lu, Shanbhag, Al Hassanieh) proposes three key innovations: filter-based rendering to suppress irrelevant signals, a physics-based RF volumetric rendering pipeline, and a novel lensless sampling strategy with alpha blending that makes training feasible.
By learning signed distance functions, reflectiveness, and signal power through MLPs and trainable parameters, GeRaF achieves millimeter-level geometry reconstruction from real-world RF measurements. Accepted as a Spotlight paper at NeurIPS 2025, this work opens the door to RF-based 3D sensing that works in complete darkness, through walls, and under heavy smoke—scenarios where LiDAR and cameras fail. The method is validated on real-world datasets, showing high fidelity compared to ground truth.
- First neural implicit approach for 3D geometry reconstruction from radio frequency signals
- Introduces filter-based rendering and lensless sampling to handle noise and specular reflections
- Achieves millimeter-level accuracy in real-world settings, accepted as NeurIPS 2025 Spotlight
Why It Matters
Enables 3D sensing through walls and occlusions, unlocking new robotics and autonomous navigation use cases.