Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..

GeRaF: Neural Geometry Reconstruction from Radio Frequency Signals

Authors: Jiachen Lu, Hailan Shanbhag, Haitham Al Hassanieh

NeurIPS 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate our system using a 77 GHz mm Wave radar mounted on a robotic arm, scanning a variety of real-world objects. Our results demonstrate that Ge Ra F takes a significant first step toward accurate 3D reconstruction from RF signals, outperforming currently adopted methods in both reconstruction quality and robustness to experimental noise.
Researcher Affiliation Academia Jiachen Lu*, Hailan Shanbhag*, Haitham Al Hassanieh École Polytechnique Fédérale de Lausanne (EPFL)
Pseudocode Yes Algorithm 1 Differentiable Matched Filter Algorithm Algorithm 2 Signal Tracing
Open Source Code No We plan to release the training code, experiment code and dataset in the future.
Open Datasets No As there are no publicly available radar datasets for near range imaging, we collected our own dataset using a TI 1843BOOST mm Wave radar [53] with a DCA1000EVM for raw data collection, attached to a Franka Research 3 controlled via the Franka Py library [64].
Dataset Splits Yes To evaluate this, we split the available views into training and test sets. Ge Ra F is trained using only the training views and evaluated on the held-out test views. To assess the quality of the synthesized matched filter images, we compute the Peak Signal-to-Noise Ratio (PSNR) between the rendered outputs and the ground truth matched filter images. Figure 13 illustrates the detailed data split.
Hardware Specification Yes We trained our model for 50,000 iterations over 32 hours on a single NVIDIA H100 GPU
Software Dependencies No The paper mentions using 'mm Detection3D [10] as the code base' but does not specify a version number. Other cited works like Nerf and NeuS are mentioned for their code base influence, but again, no specific versions of these or other key software components are provided for the authors' implementation.
Experiment Setup Yes For the SDF Network, we use an MLP with 8 layers and a hidden dimension of 256. We apply sinusoidal positional encoding with 10 frequency levels as input. The Reflectivity Network is implemented as an MLP with 4 layers and a hidden dimension of 256. The signal power prediction is implemented as a single trainable parameter. We trained our model for 50,000 iterations over 32 hours on a single NVIDIA H100 GPU... We used an initial learning rate of 1 10 3, but due to the sparsity of the input, the learning rate for the SDF Network was reduced to 1 10 4. Training was performed using the Adam W optimizer with cosine annealing learning rate scheduling.