Research map / 3D vision / 3D generation
Radio Frequency Ray Tracing with Neural Object Representation for Enhanced RF Modeling
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces RFScape, an object-centric neural SDF-plus-material framework integrated into RF ray tracing that predicts WiFi, BLE, 60 GHz and mmWave channels from sparse measurements.
Problem
Conventional RF ray tracing suffers a sim-to-real gap due to crude mesh geometry and homogeneous material assumptions, while neural scene-level RF models like NeRF2 need dense sampling and full retraining when scenes change.
Approach
Learn a per-object neural SDF (structure) and a directional neural attenuation model (material, including interior points) and plug them into sphere-tracing RF ray tracing, so objects can be individually trained, edited, and recomposed in a scene.
Key results
- Object-level: median error 2.9 dB for RFScape vs 15.7 dB for ray tracing on scanned meshes (about 13 dB improvement).
- Small-data WiFi CSI: median SNR 15.6 dB, outperforming NeRF2 by 5.3 dB at 1.25 samples/sq ft.
- 60 GHz WiGig CSI: RFScape reaches 16.2 dB SNR vs 9.08 dB for NeRF2.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use neural object representations instead of scanned meshes for RF ray tracing | RFScape · Object-level RF characterization | median error: 15.7 dB -> 2.9 dB | Helped |
| Use RFScape rather than NeRF2 | RF channel prediction · BLE RSSI, small training dataset | RSSI error: 1.2 dB better | Helped |
| Use RFScape rather than NeRF2 | RF channel prediction · 2.4 GHz WiFi CSI, small training dataset | median SNR: 5.3 dB | Helped |
| Use RFScape rather than NeRF2 | RF channel prediction · RFID spectrum, small training dataset | SSIM: 0.4498 vs 0.5061 | Hurt |
| Use RFScape rather than NeRF2 | RF channel prediction · mmWave CSI | SNR: 9.08 dB -> 16.2 dB | Helped |
| Adapt the scene using visual information | RFScape · Dynamic scene adaptation | median RSSI error: 2.9 dB | Helped |
| Adapt the scene using 3 to 5 new data points | RFScape · Dynamic scene adaptation | median RSSI error: 3.2 dB | Helped |
Surprising findings
- NeRF2 failed to learn from FMCW radar signals, so it was excluded from object-level comparison.
Limitations
- NeRF2 baseline reproduced from public code; comparison limited to a few wireless protocols and small set of objects
- Object-level evaluation uses only three objects on a rotation stage in an absorber-isolated setup
- Material parameters for the mesh ray tracing baseline taken from literature
Experimental setup
Datasets: NeRF2 dataset (2.4 GHz WiFi CSI, BLE RSSI, RFID spectrum) self-collected 60GHz WiGig CSI self-collected 5GHz WiFi RSSI FMCW mmWave radar object responses
Benchmarks: object-level RF characterization (mmWave radar, rotation stage) room-level channel prediction (WiFi CSI, BLE RSSI, RFID, 60GHz WiGig CSI) dynamics adaptation
Metrics: RSSI error (dB) SNR (dB) SSIM
Baselines: NeRF2 conventional mesh ray tracing (Polycam scanned mesh)
Builds on: NeRF NeRF2 signed distance functions sphere tracing WiNeRT
Closest papers by meaning
- Physics Unrolled Neural Operator for Wireless Field Modeling (NeurIPS 2026 · similarity 0.86)
- RxGS: Receiver-Generalizable 3D Gaussian Splatting for Radio-Frequency Data Synthesis (NeurIPS 2026 · similarity 0.86)
- GeRaF: Neural Geometry Reconstruction from Radio Frequency Signals (NeurIPS 2025 · similarity 0.84)
- Differentiable and Learnable Wireless Simulation with Geometric Transformers (ICLR 2025 · similarity 0.84)
- Can NeRFs "See" without Cameras? (NeurIPS 2025 · similarity 0.84)
- GSRF: Complex-Valued 3D Gaussian Splatting for Efficient Radio-Frequency Data Synthesis (NeurIPS 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.