- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper studies the synthesis of dense range–azimuth–Doppler (RAD) radar cubes from LiDAR geometry. The central idea is "support-to-response" generation. A calibrated LiDAR occupancy grid and the radial velocities derived from target motion are first folded and quantized according to the sensor's unambiguous velocity interval, yielding a "clean" RAD endpoint X₁ whose RA and DA marginals are exactly consistent. A paired Schrödinger bridge in the I2SB form then performs stochastic transport from X₁ to the measured RAD cube X₀. The predictor is a four-stage 3D U-Net with RA/DA conditioning pyramids and a proposed Radar Spread Adapter (RSA), which applies separable multi-scale depthwise convolutions along the three physical axes to capture the non-interchangeability of range, azimuth and Doppler. In addition to the bridge loss, training includes two statistical constraints: source-anchored spread-profile consistency (SPC) and off-support multi-axis texture consistency (MATC).
A single cross-domain model is trained on four radar domains (K-Radar, RaDelft, ColoRadar cascade and ColoRadar single-chip) and is reported to outperform the baselines. The appendix also reports zero-shot cross-sensor transfer on an in-house indoor sequence collected with a TI AWR1843 radar and a Hesai JT128 LiDAR.
Critical strengths
-
I like the modeling idea and its physical grounding: the paper cleanly separates what LiDAR can determine (geometric support, Doppler folding induced by ego-motion) from what a generative model must learn (reflectivity, multipath, clutter, point spread), and uses a maximum-entropy coupling to ensure that the endpoint introduces no unobserved range–Doppler correlations.
-
Training a single model across four heterogeneous radar domains, distinguished by sensor embeddings, together with zero-shot transfer to an in-house TI AWR1843 indoor sequence, provides fairly convincing evidence of generalization.
Critical weaknesses
For a radar data-generation paper, the most important question is whether the proposed method actually produces better results. The results reported so far do not establish this convincingly:
-
The comparison in Table 1 of the main text is not fair, because the proposed method is trained for longer. According to Appendix G.1, the DDPM and flow-matching baselines are trained on about 230k samples, whereas the RadarSimRAD entry in Table 1 is the final checkpoint trained on about 350k samples, i.e., 50% more. The appendix does contain a compute-matched version (10k steps), and its voxel fidelity is not better than flow matching (Appendix Tables 13 and 7): its Corr is below both DDPM and flow, and its MAE is the worst of all methods. In other words, at equal training budget the method leads only on "response-formation" metrics such as AGLRE, not on voxel fidelity. The statements in the abstract and Section 4.2 ("largest and most consistent gains … while maintaining competitive voxel-level fidelity") are not consistent with the appendix. I suggest reporting the compute-matched comparison directly in the main text and moving the final checkpoint to the supplementary material.
-
The AGLRE metric and the MATC training loss are essentially the same statistic. AGLRE measures the log error of the off-support per-axis gradient RMS ratios, while MATC directly minimizes the log error of the off-support per-axis lag-1/2/4 finite-difference energies; the baselines do not use this loss. Consequently, the claim in the abstract of "reducing axis-gradient error by 55.8%" carries little weight, since it mainly shows that adding a loss improves the metric that corresponds to that loss. The appendix does provide loss-independent spectral diagnostics (PSD-W improves by 32–39%), and these should be the headline numbers in the main text. In addition, I could not find how the 55.8% and 38.1% figures in the abstract are computed anywhere in the main text or appendix; please clarify.
-
On K-Radar the method is actually behind the baselines, and its behavior is unstable. Appendix Table 6 shows that on the K-Radar domain RadarSimRAD has lower Corr than flow, lower Peak-F1 than DDPM, and worse MAE. In the per-sequence Peak-F1 of Table 9, DDPM/flow reach 0.21–0.33 on the three K-Radar sequences, whereas the "w/o MATC" and "w/o RSA" ablations almost collapse, the 10k version reaches only 0.07–0.23, and only the 14.5k version recovers to 0.25–0.29. The AGLRE in Table 10 shows the same pattern on K-Radar: the 10k version is at 0.85–1.02, nearly identical to the ablations, and the 14.5k version suddenly drops to 0.19–0.27. Such a jump, together with the collapse of the ablations on a single domain, looks more like training instability or non-convergence on that domain than a contribution of the proposed mechanisms.
-
The fourth claimed contribution, "integration with CARLA, LIBERO and RoboTwin", is not supported by any experiment, figure or description.
For these reasons, my current assessment of the paper is negative. I would be glad to raise my score if the authors can address these issues convincingly during the rebuttal.
Questions for authors
See Weaknesses.