acceptodds
Withdrawn from ICLR 2027

RadarSimRAD: From Clean Support to Measured Response: Sensor-Conditioned Paired Schr\"odinger Bridges

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Doppler radar provides complementary geometric and motion cues for autonomous perception, yet large-scale radar data remain costly to acquire at scale. LiDAR offers an attractive source for radar synthesis through accurate geometric support and mature acquisition and annotation pipelines; however, existing simulators either rely on expensive physics-based modeling or use generic learned generation that underexploits source-determined structure and leaves radar-specific measurement formation largely implicit. This limitation is particularly severe for range–azimuth–Doppler (RAD) data, where range, azimuth, and Doppler exhibit distinct response statistics and Doppler ambiguity, often leading generic generators to over-smooth weak responses or hallucinate unsupported structures. We introduce RadarSimRAD, a sensor-conditioned paired Schrödinger bridge that learns stochastic transport from an alias-aware clean geometric support endpoint to measured radar response. To model the heterogeneous formation process, we further propose a Radar Spread Adapter to explicitly capture sensor-dependent, non-exchangeable response formation across range, azimuth, and Doppler. Extensive experiments demonstrate that RadarSimRAD substantially improves radar-response fidelity over baselines, reducing axis-gradient error by 55.8% and spectral distance by 38.1% over baselines, while exhibiting zero-shot cross-sensor transfer on independently collected real-world measurements. Generated RAD further improves multiple downstream tasks under limited measured-data availability, demonstrating its practical value for radar learning in data-scarce regimes.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Radar Generation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_tDw7OpenReview ↗
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
  1. 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.

  2. 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:

  1. 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.

  2. 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.

  3. 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.

  4. 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.

Reviewer_FphjOpenReview ↗
Rating
2
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a sensor-conditioned bridge model for generating dense radar observations from LiDAR-derived support. To retain geometric and motion priors, the authors construct an alias-aware RAD endpoint and learn its translation to measured responses. The authors also introduce a Radar Spread Adapter and additional training losses to make the generated radar signals more realistic. Experiments across several radar domains show gains in response fidelity and downstream tasks.

Critical strengths
  1. The paper is clearly written and supported by numerous well-presented figures and tables. The work represents substantial effort, with extensive experiments across radar domains and downstream tasks, including a zero-shot evaluation using a custom hardware setup.
  2. The explicit RAD endpoint and directional response modeling offer some methodological value beyond a generic conditional generator. The planned plugins for CARLA, LIBERO, and RoboTwin could also be useful resources for the community.
  3. The convolution and loss designs are well motivated, accounting for differences between the radar axes and addressing both local echoes and background detail.
Critical weaknesses
  1. The conceptual novelty appears overstated. L2RLDB already formulates dense radar generation from LiDAR as a "source-to-target" diffusion bridge, closely resembling the proposed "support-to-response" framing, while SA-Radar constructs reflection support before generating radar responses; the endpoint constructed from occupancy seems to offer a more limited methodological advance.
  2. MATC shows clear improvements, but RSA performs worse on several metrics under the same training budget, which limits my confidence in its effectiveness. The contribution of SPC is also difficult to assess without a separate ablation.
  3. The L2RLDB comparison uses the authors' own reimplementation, with visibly poor reconstructions that differ markedly from the examples in the original paper. I therefore have reservations about the validity of this comparison and the claimed superiority over L2RLDB.
Questions for authors
  1. Are any RD examples available? I would also appreciate some clarification on the accuracy and noise level of the velocity estimates derived from LiDAR.
  2. I would appreciate a more complete ablation analysis, including the separate contribution of SPC and an explanation of why RSA improves some metrics but worsens others under the same training budget.
Reviewer_mrT9OpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes RadarSimRAD, a conditional generative model that outputs radar range-azimuth-Doppler (RAD) tensors from LiDAR pseudo-RAD “scene support.” This initializes a diffusion process conditioned on the scene and a learned sensor identity embedding, and the authors argue this decouples scene geometry from sensor-specific measurement formation while also addresses Doppler phase-wrapping ambiguity. The model is trained on four public radar datasets and evaluated against generative baselines (including flow matching) using RAD volumetric correlation, RA correlation, Peak-F1, and Spread. The authors also report ‘zero-shot cross-sensor transfer’ to self-collected data from a TI AWR1843.

Critical strengths
  1. The motivation to model scene content (via LiDAR initialization) and sensor characteristics independently is well-founded and physically inspired. The authors identify the right open problems in radar data synthesis and also a reasonable direction for addressing them.
  2. The paper is clearly written, and the qualitative results are visually convincing, as does the flow matching baseline qualitative results. However my concerns below limit its significance.
Critical weaknesses
  1. Despite the paper’s framing, radar physics is left almost entirely implicit. The abstract criticizes prior work for leaving radar measurement formation “largely implicit” but the proposed method leaves nearly all of this to the generator: the point spread function along range, azimuth, and Doppler; propagation; radar cross-section; reflection, absorption, transmission, and diffraction; multipath clutter; polarization; and the antenna beam pattern. Much of this is well understood and, in the case of the PSF, fully deterministic given the ADC-to-RAD processing chain (FFT sizes, choice of window such as Hann). That processing chain varies across datasets independently of the sensor hardware, yet the method folds it into the sensor identity s, which is not physically correct. Why not warm-start the model with known physics prior instead of learning it from scratch? Without this, it is difficult to distinguish the contribution from prior diffusion-based LiDAR-to-radar approaches beyond applying a stronger generative backbone with more data (four datasets) and more compute (8× L40S GPUs).
  2. The zero-shot cross-sensor transfer claim is not supported. Sensor conditioning as described in the paper consists only of a sensor identity s. It is not clear how a categorical identity can generalize to a radar unseen in training, which is the true definition of zero-shot cross-sensor transfer. Both the ColoRadar training data and the self-collected data used for the transfer experiment uses a TI AWR1843, so the physical tx/rx antenna layout, virtual array aperture, and essentially the antenna radiation beam pattern (up to manufacturing tolerance) are the same. Appendix H.1 concedes that inference reuses a sensor identity seen during training. This is in-distribution transfer, not zero-shot.
  3. Gains over the strongest baseline are marginal. The flow matching baseline shares the “3D predictor, RA/DA encoders, sensor conditioning, preprocessing, and data splits” with RadarSimRAD, so it effectively isolates the choice of generative formulation. Compared to this baseline, RadarSimRAD reports RAD Corr 0.7077 vs. 0.702, Peak-F1 0.3104 vs. 0.2884, Spread 0.5423 vs. 0.5606 , and RA Corr 0.6915 vs. 0.6683. These differences are small, and no variance across seeds or confidence intervals is reported. The experiments therefore do not show that the paper’s specific formulation offers a meaningful advantage over a standard conditional generative model trained on the same setup, whose central idea is LiDAR initialization as per Figure 1.
Questions for authors
  1. How does the method differ conceptually and empirically from existing diffusion-based LiDAR-to-radar generation? And also from the implicit prior methods referenced in the abstract?
  2. Why did the authors choose not to incorporate explicit physics priors for the phenomena currently learned from scratch (see Weakness 1 for full list)?
  3. The paper has no limitations section despite clear limitations (implicit physics, closed-set sensor conditioning, dependence on dataset-specific processing). Will the authors commit to adding one?
  4. Can the authors expand the related work to include radar synthesis methods that transfer to sensors unseen during training?
  5. Can you please cite prior work that uses the RA/RAD correlation metrics in the metrics paragraph?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.