{"id":"52ba6616-fe97-4533-99bb-c3dcee98dec7","arxiv_id":"2505.13905","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A 4D radar-only occupancy model trained with LiDAR pseudo-labels and height-map supervision approaches LiDAR-quality occupancy and runs at about 30 Hz.","lead":"4D-ROLLS trains a 4D radar occupancy network using LiDAR point clouds as a free teacher, turning sparse radar returns into a dense 3D occupancy map for vehicles. It is of interest because it offers a path to all-weather, real-time 3D perception without needing LiDAR or cameras at inference time.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Pseudo-labels from LiDAR both supervise and evaluate the model, so the reported quantitative gains may only show better imitation of LiDAR ray-casting, not true all-weather occupancy; Section IV.B concedes radar-penetrated objects are suppressed.","rationale":"The reader's weakest assumption already identifies the Section III.A ray-casting assumption, and the reader's rationale notes that the headline accuracy is evaluated against the same LiDAR-derived occupancy that supervises training. My stress-test converges on that point and sharpens it: the evaluation is not merely affected by label noise, it is self-referential for the all-weather portion of the claim. This is a real limitation, but it is explicitly acknowledged in Section IV.B, and the relative comparison against radar-adapted LiDAR baselines still supports the narrower claim that 4D-ROLLS fits LiDAR-style occupancy queries better than those baselines. The paper also provides qualitative cross-dataset and smoke experiments and promises code, which are useful supporting evidence. Because the concern can be addressed by an independent held-out label evaluation rather than by a correction to the method, the appropriate verdict remains conditional rather than rejection. No change to the reader's verdict is needed.","tokens_in":10668,"tokens_out":6421,"duration_ms":67403,"concrete_test":"On a held-out subset of MSC and NTU frames, construct test occupancy labels independently of the Section III.A routine, for example by temporally accumulating multiple LiDAR scans into a denser point cloud and ray-casting from those merged scans, or by manually annotating voxels from synchronized RGB images. Recompute Table I (CD, L2, AR) for Ours-stage2, ALSO-R, and UNO-R against these independent labels. If the relative gains persist, the circular-evaluation concern is resolved; if the margin shrinks or reverses, the reported comparison is an artifact of using the same LiDAR pseudo-labels for training and testing.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The evidence for the headline claim is circular. Section III.A generates occupancy queries from each LiDAR point by assuming all voxels along the sensor-to-point ray are free and placing R+ = p + r u just behind the hit as occupied. These queries are the training target L_occ in Eq. 3, and on MSC and NTU they are also the reference for the CD, L2, AR, and NFCD metrics in Tables I and II. Ours-stage2 therefore reports how well the radar network reproduces the LiDAR-derived labels compared with the radar-adapted baselines; it does not measure occupancy accuracy in regions where LiDAR is blind. The all-weather scenario is precisely where this matters: in fog and smoke the LiDAR supervision is unreliable or absent, and in clear weather radar can detect objects through partial occlusion or multipath that the ray-casting labels mark as free. The authors explicitly concede this in Section IV.B: 'in extreme cases, radar may capture objects after penetration, while the LiDAR supervision signal completely ignores them, causing the network to learn a biased world model.' Without independent ground-truth occupancy, the paper cannot establish that the model 'preserves all-weather sensing' rather than inheriting LiDAR's specific errors.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes 4D-ROLLS, a 4D radar occupancy estimation method trained with weak supervision from LiDAR. The method generates pseudo-labels consisting of occupancy queries (occupied and free points placed along LiDAR rays) and a LiDAR height map, and uses these in a multi-term loss (Eq. 3) with a sparse TPV encoder and a dense decoder. A second fine-tuning stage aligns the radar model with a self-supervised LiDAR occupancy model. Experiments on the MSC and NTU datasets compare against LiDAR occupancy methods adapted to radar input, reporting lower Chamfer distance, L2 error, and relative L2 error, qualitative robustness under smoke, cross-dataset generalization, transfer to BEV segmentation and point cloud occupancy prediction, and inference at about 30 Hz on a 4060 GPU.","tokens_in":10869,"tokens_out":4567,"duration_ms":42436,"significance":"If validated, the paper addresses a practically important gap: 4D radar point clouds are sparser and noisier than LiDAR, and no open-source 4D radar point-cloud occupancy method exists. The proposed training recipe is simple, the network is lightweight, and the authors provide code. The cross-dataset test and downstream-task transfer are useful evidence of representation quality. However, the central quantitative evaluation is weakened by the fact that the LiDAR-derived pseudo-labels are used both as training targets and as evaluation reference, so the reported gains may measure imitation of LiDAR ray casting rather than independent occupancy accuracy. The paper's all-weather claim is also not quantitatively supported, and the authors themselves concede a bias in the learned world model for radar-penetrated objects. The contribution is valuable as a LiDAR-supervised distillation approach, but the claims need to be reframed or supported by independent ground truth.","major_comments":[{"comment":"The occupancy queries R+ and R- are generated by ray casting from LiDAR points, and these same queries are used as the training target L_occ in Eq. (3) and as the reference for the CD, NFCD, AR, and L2 metrics in Tables I and II. Consequently, the reported gains of Ours-stage1 and Ours-stage2 measure how well the radar network reproduces the LiDAR-derived pseudo-labels, not occupancy accuracy against independent ground truth. This is especially problematic for the all-weather claim: in fog or smoke the LiDAR pseudo-labels are unreliable, and in clear weather radar can detect objects that the ray-casting assumption marks as free. The paper should either add an evaluation with independent occupancy ground truth (e.g., manually labeled voxels or geometric annotations not derived from the same LiDAR scans) or substantially soften the claims of all-weather occupancy accuracy.","section":"Section III.A and Tables I/II"},{"comment":"The text concedes that radar may capture objects after penetration while the LiDAR supervision signal completely ignores them, 'causing the network to learn a biased world model.' This admission directly qualifies the central claim that the method preserves all-weather sensing. The only degraded-weather result is qualitative (Fig. 4(a)), where LiDAR cannot provide a reference; thus the quantitative tables do not support the abstract's assertion of robustness in degraded environments. Please either provide quantitative evaluation in degraded conditions with independent labels or reframe the contribution as LiDAR-style occupancy distillation rather than all-weather occupancy estimation.","section":"Section IV.B, last paragraph"},{"comment":"The baseline adaptation protocol is underspecified. It is not stated whether ALSO-R and UNO-R are trained with the same LiDAR-generated occupancy queries and height maps as 4D-ROLLS, what loss functions are used, or how hyperparameters are chosen. Also, the LiDAR-input methods (ALSO-L, UNO-L) are labeled as references but are not upper bounds for a radar-only system; please clarify their role. A fair comparison requires identical supervision and evaluation protocol for all radar-input models.","section":"Section IV.B and Section IV.A"}],"minor_comments":[{"comment":"The text contains typos: 'Jeston orin' should be 'Jetson Orin' and 'repectively' should be 'respectively'; these appear in Section IV.A and in the footnotes of Tables I and II.","section":"Section IV.A"},{"comment":"Section IV.A lists NFCD as a metric, but Table I reports only CD, AR, and L2; either add NFCD to Table I or remove it from the metric list for consistency.","section":"Section IV.A and Table I"},{"comment":"The equations defining R+ and R- are not numbered; number them for clarity and ease of reference.","section":"Section III.A"},{"comment":"The term 'relative L2 error (AR)' is not standard and the abbreviation AR is not defined; please rename or define it explicitly.","section":"Section IV.A"},{"comment":"The caption uses '(cross-dataset training)' twice and does not specify which panels correspond to same-dataset versus cross-dataset testing; please annotate the panels clearly.","section":"Fig. 4 caption"},{"comment":"Reference [1] appears to cite 'COLMAP' but the title describes an occupancy grid mapping framework; please verify the citation and correct the author or title if needed.","section":"References"},{"comment":"The statement that the MSC dataset comprises 90,000 frames is not connected to the per-sequence results in Table I; clarify whether training uses all frames or only the listed sequences.","section":"Section IV.A"}],"recommendation":"major_revision","confidential_remarks":"The circular evaluation is a serious concern, but the authors explicitly acknowledge the penetration limitation, and the core training pipeline is sensible and reproducible. I would encourage the editor to invite a revision that adds independent evaluation or reframes the claims as LiDAR-style distillation. I do not see evidence of deliberate misuse; the issue is the scarcity of independent occupancy benchmarks in this area. The paper's fit for the journal is acceptable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a sensible, clearly written engineering contribution with a real novelty claim that mostly checks out, but the quantitative evaluation overstates what it proves because the same LiDAR pseudo-labels supervise and evaluate the model. I side with the reader's conditional verdict, not with the abstract's confidence.\n\nWhat is actually new: 4D-ROLLS is the first 4D radar occupancy method trained from LiDAR pseudo-labels on point cloud input, bypassing the raw-tensor requirement of Radarocc and the inference-time LiDAR/camera need of Occfusion and LiCRocc. The two-stage fine-tuning against a self-supervised LiDAR occupancy model is a reasonable way to prune false positives. The cross-dataset test (MSC to NTU) is a good sign, and 30 Hz on a 4060 GPU makes it genuinely real-time. The paper is also well organized; the method is described clearly enough to reproduce once code lands.\n\nThe soft spots are real but not disqualifying. Most importantly, the headline metrics—CD, NFCD, AR, L2—are computed against the same LiDAR-derived occupancy queries that generate the training signal in stage 1. So Tables I and II measure how well the radar network imitates the LiDAR teacher compared to the adapted baselines, not how accurate occupancy is where LiDAR fails. The authors are upfront about this in Section IV.B, where they concede that radar-penetrated objects are suppressed because the LiDAR supervision ignores them. That concession undercuts the all-weather claim: the only evidence of fog/smoke robustness is qualitative. Also absent are error bars, per-seed variance, and the actual values for key hyperparameters like r and the loss weights. The downstream BEV/occupancy results are built on pseudo-labels from another model (HRNet), so they are suggestive, not proof. None of these are fatal; they are addressable in revision.\n\nI would take this paper seriously as a referee. The idea is useful, the writing is honest about the main limitation, and the approach is likely to become a baseline for radar-only occupancy. The revision should add a small set of manually annotated or independently measured occupancy in clear weather, a controlled degraded-weather test with camera/visual reference, and reported variance. With those, I'd support publication; without them, the claim about all-weather sensing remains an interesting hypothesis.","headline":"A useful and clearly presented weakly supervised radar occupancy pipeline; the novelty holds up, but the evaluation is partly circular and the all-weather claim is thinner than the abstract suggests.","tokens_in":11469,"tokens_out":2709,"would_cite":true,"duration_ms":26018,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"4D-ROLLS claims that a radar-only network can estimate 3D occupancy at near-LiDAR accuracy, trained with no manual labels by using the LiDAR point cloud purely as a training-time teacher.","keywords":["4D radar","occupancy estimation","weakly supervised learning","LiDAR supervision","pseudo-label generation","tri-perspective view","autonomous driving","all-weather perception"],"falsifier":"Build a small labeled evaluation set of scenes with objects thinner than the occupancy voxel (0.4 m) or with surfaces radar can penetrate, such as foliage, wire mesh, or thin plastic panels, and compare the trained model's occupancy map against a dense ground-truth scan. If the model marks the space behind these surfaces as free even where genuine radar returns exist behind them, the LiDAR ray-casting assumption has propagated into the predictions; measuring a systematic occupancy error on such objects that grows with their transparency would settle whether the supervision bias is real.","tokens_in":10423,"feed_emoji":"📡","tokens_out":9656,"duration_ms":77623,"temperature":0.7,"pith_summary":"This paper tries to establish that 4D radar point clouds alone are enough to produce dense 3D occupancy maps that rival LiDAR-based ones, provided the radar network is trained under LiDAR's supervision. Because LiDAR and cameras fail in rain, snow, fog, and smoke while 4D radar keeps sensing, a radar-only occupancy model would let an autonomous vehicle keep a reliable picture of occupied and free space in exactly the conditions where current sensors degrade. The paper's move is to turn an unlabeled LiDAR scan into two free supervisory signals: occupied and free query points sampled along each LiDAR ray, and a per-column height map. The radar network is trained on those signals, then fine-tuned against a self-supervised LiDAR occupancy model to prune spurious detections. The reported result is that this radar model beats LiDAR-derived occupancy baselines re-targeted to radar input, transfers across datasets, and runs at about 30 Hz.","feed_headline":"Radar-only occupancy maps hit LiDAR quality with no manual labels","feed_subtitle":"A radar-only network keeps mapping the world in fog and smoke where LiDAR fails — at 30 Hz on a laptop GPU.","key_machinery":"The load-bearing object is the ray-sampled occupancy query pair. For a LiDAR point $p_i$ and unit ray direction $\\mathbf{u}$ from the sensor to the point, the occupied query $R^+_i = p_i + r\\mathbf{u}$ and the free query $R^-_i = p_i - r\\mathbf{u}$ encode the assumption that the sensor-to-surface ray is empty and the cell just beyond the surface is filled; the whole training signal is a set of such pairs. The second mechanism is the LiDAR height map, a projection in which each XY grid cell stores the maximum Z value of the LiDAR points above it, supervised at two network depths by a masked MSE loss computed only where radar points exist. These two pseudo-label types drive a sparse tri-perspective-view encoder (three perpendicular 2D planes: bird's-eye, front, side) feeding a dense UNO-style decoder and an occupancy head, with total loss $L = \\omega_1 L^1_{\\mathrm{height}} + \\omega_2 L^2_{\\mathrm{height}} + \\omega_3 L_{\\mathrm{occ}}$, followed by stage-2 fine-tuning against a LiDAR self-supervised occupancy model.","core_discovery":"The central claim, stated as the authors would state it, is that the gap between sparse, noisy 4D radar and LiDAR-grade occupancy is mostly a supervision gap, not a sensing gap. For each LiDAR point $p_i$ the method labels the sample $R^+_i = p_i + r\\mathbf{u}$ (just behind the surface) as occupied and $R^-_i = p_i - r\\mathbf{u}$ (toward the sensor) as free, converting a raw LiDAR scan into a dense set of occupancy queries for free. A second signal, the LiDAR height map recording the maximum Z value in each XY voxel, is used to pin down the vertical axis where radar's height resolution is weakest. A two-stage procedure then trains the radar network: stage 1 fits the occupancy field and height maps to the pseudo-labels, and stage 2 fine-tunes it against a self-supervised LiDAR occupancy model, removing false occupied regions that radar's sparse coverage produces. The authors claim this yields radar-only occupancy estimation that significantly outperforms the LiDAR-occupancy baselines adapted to radar input on all key metrics, remains stable in fog and smoke where LiDAR supervision itself degrades, and generalizes across datasets; they also acknowledge the trade-off that radar returns from behind objects, which LiDAR supervision ignores, can bias the learned model.","pith_inferences":["If the supervision signal is the bottleneck, then upgrading the teacher — a denser LiDAR, or LiDAR fused with cameras — should directly raise radar occupancy accuracy in normal weather; this is a testable consequence the paper does not run.","The height-map branch specifically compensates radar's weak vertical resolution; ablating it would quantify how much of the reported gain is Z-axis correction versus occupancy-query learning.","Radar returns that persist behind LiDAR-defined surfaces are treated as errors to be pruned, but they could instead be harvested as evidence of transparent or thin objects, turning the known penetration bias into a material-boundary signal.","The reported cross-dataset tilt from the NTU sensor mount suggests that applying the method to a new radar first requires precise extrinsic calibration; a small misalignment could otherwise be baked into the learned occupancy field."],"forward_implications":["Autonomous vehicles could carry radar as the sole occupancy sensor at inference time, remaining functional in fog, smoke, rain, and snow where LiDAR and cameras degrade.","Any LiDAR-equipped vehicle produces free, unlimited training labels for radar occupancy, so the method scales with data collection rather than annotation budgets.","The stage-2 fine-tuning recipe — train a denser sensor's self-supervised occupancy model and use it to prune a sparser sensor's false predictions — transfers geometric fidelity across sensor types without manual labels.","The learned occupancy representation serves as a pretrained backbone for BEV segmentation and point-cloud occupancy prediction with only simple fine-tuning, indicating it captures general scene structure.","At about 33 ms per frame on a mid-range GPU, the lightweight network meets real-time constraints for robotic navigation and planning."],"supporting_citations":[{"why":"Provides the self-supervised occupancy query generation scheme (occupied/free sampling along rays) that the pseudo-label pipeline is built on.","marker":"[9]"},{"why":"Supplies the dense decoder architecture, the self-supervised LiDAR occupancy model used for stage-2 fine-tuning, and the evaluation metrics.","marker":"[26]"},{"why":"Provides the tri-perspective view sparse encoder backbone used to extract features from the sparse 4D radar point cloud.","marker":"[16]"},{"why":"Supplies the height-projection idea behind the LiDAR height map used as auxiliary supervision.","marker":"[27]"},{"why":"The MSC-Rad4R dataset, the main source of training and testing frames.","marker":"[23]"},{"why":"The NTU4DRadLM dataset, used for additional evaluation and cross-dataset generalization tests.","marker":"[22]"},{"why":"SemanticKITTI point labels used to supervise the BEV segmentation downstream task.","marker":"[34]"},{"why":"HRNet used to generate image-derived pseudo-labels for the 3D occupancy prediction downstream task.","marker":"[33]"}],"fun_headline_variants":["Radar-only occupancy hits LiDAR quality, no manual labels","Radar learns occupancy from LiDAR pseudo-labels, no annotations","Radar occupancy works in smoke and fog, 30 Hz with no labels","First radar-only occupancy model matches LiDAR via pseudo-labels","Radar-only occupancy rivals LiDAR in degraded environments"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole training signal rests on assuming that each LiDAR point lies exactly on an object surface, so every voxel between sensor and point is empty and the voxel just behind the point is occupied; if the object is thinner than the offset $r$, partially hidden, or mis-measured by LiDAR, the pseudo-labels themselves are wrong and the network faithfully learns those errors.","fun_headline_variants_meta":{"raw":{"variants":["Radar-only occupancy hits LiDAR quality, no manual labels","Radar learns occupancy from LiDAR pseudo-labels, no annotations","Radar occupancy works in smoke and fog, 30 Hz with no labels","First radar-only occupancy model matches LiDAR via pseudo-labels","Radar-only occupancy rivals LiDAR in degraded environments"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000637,"raw_usage":{"total_tokens":3005,"prompt_tokens":1081,"completion_tokens":1924,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":697,"completion_tokens_details":{"reasoning_tokens":1836}},"tokens_in":697,"tokens_out":1924,"duration_ms":15801,"temperature":1.0,"reasoning_tokens":1836,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:08:07.604590+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Build a small labeled evaluation set of scenes with objects thinner than the occupancy voxel (0.4 m) or with surfaces radar can penetrate, such as foliage, wire mesh, or thin plastic panels, and compare the trained model's occupancy map against a dense ground-truth scan. If the model marks the space behind these surfaces as free even where genuine radar returns exist behind them, the LiDAR ray-casting assumption has propagated into the predictions; measuring a systematic occupancy error on such objects that grows with their transparency would settle whether the supervision bias is real.","supporting_citations":[{"cited_title":"Also: Automotive lidar self-supervision by occupancy estimation,","cited_arxiv_id":null,"evidence_quote":"Provides the self-supervised occupancy query generation scheme (occupied/free sampling along rays) that the pseudo-label pipeline is built on."},{"cited_title":"Uno: Unsupervised occupancy fields for perception and forecasting,","cited_arxiv_id":null,"evidence_quote":"Supplies the dense decoder architecture, the self-supervised LiDAR occupancy model used for stage-2 fine-tuning, and the evaluation metrics."},{"cited_title":"Msc-rad4r: Ros-based automotive dataset with 4d radar,","cited_arxiv_id":null,"evidence_quote":"The MSC-Rad4R dataset, the main source of training and testing frames."},{"cited_title":"Ntu4dradlm: 4d radar-centric multi-modal dataset for localization and mapping,","cited_arxiv_id":null,"evidence_quote":"The NTU4DRadLM dataset, used for additional evaluation and cross-dataset generalization tests."}],"review_version":1}