{"id":"17687a33-1b8e-4f90-978c-181e60fe0ea9","arxiv_id":"2509.07932","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":3.5,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A baseline study showing Neuralangelo reconstructs static synthetic satellite meshes with qualitative fidelity, while the claimed dynamic-scene evaluation is deferred to future work.","lead":"This paper presents a simulator for generating synthetic images of a tumbling satellite in orbit, and tests an existing 3D reconstruction method on static scenes. The advertised evaluation of four dynamic-scene reconstruction algorithms under combined spin and observer motion is promised but not delivered.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Static baseline supporting the central claim rests on a one-way C2M metric that cannot detect missing geometry, and no quantitative C2M values are reported, so 'closely match' is unsupported.","rationale":"The reader's conditional verdict is driven by the absence of dynamic experiments and by assumed-perfect poses. I agree the dynamic gap is real and central. My additional concern is that even the static baseline — the only present result — is not quantitatively supported: the one-way C2M metric is blind to missing geometry, and the paper provides no numbers. This strengthens the case for CONDITIONAL but does not move it to REJECT, because the authors explicitly acknowledge the metric limitation and the work is framed as preliminary. The concrete test would settle whether the static claim survives the metric correction; if it fails, the paper should be reframed as a simulator-and-methods note rather than a reconstruction-quality demonstration.","tokens_in":9031,"tokens_out":2775,"duration_ms":36223,"concrete_test":"Recompute the C2M comparison bidirectionally: in addition to signed distances from the reconstructed mesh to the reference mesh, compute signed distances from the reference mesh to the reconstructed mesh, or use a symmetric Chamfer distance. For each of the three models, report mean, RMS, and 95th-percentile errors, and isolate the GOES-R magnetometer region. If the reference-to-reconstruction direction shows large errors where the magnetometer is missing while the reconstruction-to-reference direction does not, the one-way metric's blind spot is confirmed and the 'critical fine details' claim fails. If both directions show small errors, the claim survives.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's only delivered result is the static Neuralangelo reconstruction, and the Abstract claims the generated meshes 'closely match the original CAD models with minimal errors and artifacts' and 'capture critical fine details.' In Section V.A, the authors compute signed C2M distances from the reconstructed mesh to the reference model, but only show example heatmaps and histograms; no mean, RMS, or percentile errors are reported. Without numeric values, 'minimal errors' is not established. More seriously, the paper itself identifies a flaw in this metric: in Figure 6, a missing magnetometer on GOES-R is not reflected in the scalar field because the nearest reference point lies on the spacecraft bus. This is a one-way distance computation — the reverse direction (reference mesh to reconstruction) would flag the missing component. Since the dynamic-scene evaluation is explicitly deferred to future work (Section VI), the static baseline is the sole evidence for the central claim. The acknowledged inability to detect omitted parts, combined with the admission that thin structures commonly fail to reconstruct, directly undercuts the claim to have captured 'critical fine details' and 'closely match[ed]' the CAD models. The concern is not that the results are necessarily poor, but that the delivered evidence cannot support the delivered claim as currently measured.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper describes a simulation framework for 3D reconstruction of an uncooperative resident space object, using Isaac Sim to render images of three CAD models (CubeSat, DESDynI, GOES-R) under Clohessy–Wiltshire relative orbits. The stated goal is to evaluate dynamic-scene NeRF variants (D-NeRF, Nerfies, TiNeuVox, HexPlane) for a tumbling target observed during a fly-around. The only delivered results, however, are static-scene reconstructions using Neuralangelo, assessed through qualitative C2M heatmaps and histograms. The dynamic evaluation is explicitly deferred to future work in Section VI. The manuscript also describes planned PSNR/SSIM/LPIPS metrics that are not yet applied.","tokens_in":9340,"tokens_out":4328,"duration_ms":51922,"significance":"If completed, the proposed evaluation would address a genuinely useful gap: the combination of target tumbling and observer fly-around creates a dual-motion scenario that existing dynamic NeRF evaluations have not covered. The Isaac Sim-based synthetic data pipeline, with ground-truth poses/depth and an inclined relative orbit designed to observe all spacecraft surfaces, is a sensible building block. As it stands, however, the only present result is a qualitative static baseline with Neuralangelo; the central advertised contribution is absent. The paper therefore reads as a progress report rather than a completed study, and the current evidence is too weak to support the abstract's claim of reconstructions that 'closely match' the CAD models.","major_comments":[{"comment":"The central claim that this study evaluates state-of-the-art dynamic-scene reconstruction algorithms is not supported by any experiment. Section VI explicitly states that the dynamic evaluation is future work. The only results are static reconstructions with Neuralangelo, while D-NeRF, Nerfies, TiNeuVox, and HexPlane are never tested. The title and abstract therefore overstate the delivered contribution. The manuscript must either include the promised dynamic experiments (with the stated PSNR/SSIM/LPIPS metrics) or be reframed as a static baseline and simulation-framework paper with appropriately revised claims.","section":"Abstract, Section II, Section VI"},{"comment":"The claim that the generated meshes 'closely match the original CAD models with minimal errors and artifacts' is not quantitatively supported. No C2M summary statistics (mean error, RMS, percentiles, or confidence intervals) are reported; only example heatmaps and histograms are shown. Moreover, the one-way C2M direction used cannot detect missing geometry, as the authors acknowledge for the missing GOES-R magnetometer in Figure 6. The reverse direction (reference mesh to reconstruction) or a volumetric/directional metric is needed to support the claim of capturing 'critical fine details.' The paper also acknowledges that thin structures commonly fail to reconstruct, which further undercuts the current wording. Please report full quantitative metrics per model, preferably normalized by model scale and accompanied by error distributions.","section":"Section V.A, Figs. 5-7"},{"comment":"Camera poses are taken as ground truth from the simulator, so the reconstruction pipeline never encounters pose estimation error. In uncooperative RPO, pose uncertainty is a primary error source and is likely to dominate reconstruction quality. This assumption is load-bearing for the static baseline because Neuralangelo requires accurate camera poses. The authors should either evaluate sensitivity to pose perturbations or clearly state this limitation and restrict the transferability claims accordingly. The same caveat applies to the planned dynamic experiments.","section":"Section IV.A, Section IV.B"},{"comment":"The Clohessy–Wiltshire state-space equation appears incorrectly typeset or dimensionally inconsistent. The A matrix rows are not aligned with the stated state vector [x,y,z,xdot,ydot,zdot], and the standard CW coupling terms (the 2n*ydot term in x-double-dot and the -2n*xdot term in y-double-dot) are not visible as written. Since this equation defines the observer trajectory used in the simulations, the authors should correct it and, ideally, validate the generated relative orbit against the analytic CW solution.","section":"Equation (2), Section IV.B"}],"minor_comments":[{"comment":"Figure referencing is inconsistent: the text says 'Example heatmaps ... can be seen in Figure 7' but Figure 7 is the histogram; the heatmap is Figure 5. Please correct the cross-references.","section":"Section V.A, Fig. 5/Fig. 7"},{"comment":"Several typographical and labeling issues: 'claasification' -> 'classification', 'a object' -> 'an object', 'results findings results' -> 'results', 'potentiality' -> 'potentially', 'DESDynl'/'DESTINL' vs. 'DESDynI', and 'Ecosytem' -> 'Ecosystem'. Reference [26] is also incomplete, missing venue and year.","section":"Throughout"},{"comment":"PSNR, SSIM, and LPIPS are image-rendering metrics and are not used in the static geometric evaluation presented in Section V. It would help to state explicitly whether these metrics will be used only for the dynamic evaluation, and how they relate to geometric accuracy.","section":"Section III, Section IV.C"},{"comment":"The description of the 45-degree inclined orbit is useful but says 'rotation of 90 degrees around o_h and a 45-degree rotation around o_r' while the DCM parameters are listed as theta=45, phi=0, psi=90. The relationship between the described rotations and the Euler-angle convention should be clarified.","section":"Section IV.A"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is closer to an extended abstract than a full paper: the dynamic evaluation promised by the title is entirely absent, and the static baseline lacks quantitative support. However, the simulation infrastructure is a reasonable foundation, and the shortcomings are addressable within the authors' stated research program rather than being fundamental errors. I would encourage an editor to request major revision, requiring either additional experiments or a substantial reframing of the contribution, before reconsidering the paper."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The thing to know: this paper's title promises dynamic-scene 3D reconstruction of a tumbling satellite, but the only delivered experiment is a static Neuralangelo baseline. The dynamic evaluation with D-NeRF, Nerfies, TiNeuVox, and HexPlane is stated as future work in Section VI. So the paper is really a simulator-and-baseline progress report, not the paper the title advertises.\n\nWhat's actually new: the Isaac Sim pipeline with Clohessy-Wiltshire relative motion, a 45-degree inclined observer orbit, and three CAD models. That is a reasonable starting point for generating synthetic RPO datasets. The authors also deserve credit for openly discussing the one-way Cloud-to-Mesh distance flaw — a missing magnetometer on GOES-R is invisible because the nearest reference point lies on the bus. That is a genuinely useful caution for anyone using C2M.\n\nThe soft spots are in the evidence, not just the framing. The abstract claims the meshes 'closely match' the CAD models with 'minimal errors,' but no C2M numbers are reported — no mean, RMS, or percentiles. The histograms alone cannot support that claim, and the authors' own admission that thin structures often vanish undercuts 'capture critical fine details.' The reverse distance (reference to reconstruction) would flag the missing magnetometer, but it isn't used. There are also technical errors: the CW state matrix in Eq. (2) has six rows and five columns, the signs in the CW equations are off, and the rotation angle labels in Eq. (5) conflict with the text. These matter if the simulator is meant to be a benchmark.\n\nThe static baseline itself is not new — prior work has run NeRF variants on static spacecraft models. The contribution here is the simulator, and that is a tool, not a scientific result. No code or data is released, so the reproducibility currently rests on the description alone.\n\nWho is this for? Someone working on synthetic data generation for space RPO, or an evaluator of reconstruction metrics, might find the pipeline and the C2M caveat useful. I wouldn't cite it in its current form. I would, however, send it to peer review: the direction is relevant, the authors are transparent about limitations, and a competent referee could push them to either add dynamic results or reframe and report quantitative numbers.\n\nBest.","headline":"Title promises dynamic reconstruction; the paper delivers only a static baseline and a simulator, with no quantitative evidence.","tokens_in":9831,"tokens_out":5057,"would_cite":false,"duration_ms":57342,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that a physics-based simulation environment plus Neuralangelo can reconstruct static satellite models whose meshes closely match their CAD references, establishing a baseline for evaluating dynamic scene reconstruction of t","keywords":["uncooperative resident space object","3D reconstruction","neural radiance fields","Neuralangelo","Clohessy-Wiltshire relative motion","Isaac Sim simulation","space debris characterization","cloud-to-mesh distance"],"falsifier":"Run the same Neuralangelo reconstruction on real imagery of an uncooperative spacecraft (e.g., from a ground-based telescope or an inspection mission) where camera poses must be estimated from the images themselves; if the mesh-to-CAD error increases sharply or thin structures disappear entirely, the simulator-only transferability claim is falsified. A simpler check within the paper's own framework: add realistic pose noise to the ground-truth camera positions and measure how C2M error degrades.","tokens_in":8963,"feed_emoji":"🛰️","tokens_out":3318,"duration_ms":43003,"temperature":0.7,"pith_summary":"The authors aim to assess how well state-of-the-art dynamic 3D reconstruction algorithms can model an uncooperative tumbling satellite observed by a moving camera spacecraft, a scenario with dual motion that has not been studied before. To support that evaluation, they built a simulation environment that generates realistic synthetic imagery of satellites under orbital lighting and relative-motion dynamics. Their delivered results so far are on static scenes: using Neuralangelo, the reconstructed meshes of three satellite models closely match the original CAD geometry, with small signed cloud-to-mesh distances and capture of fine detail, except for thin protruding structures like a magnetometer. These static results are presented as a baseline for the planned dynamic evaluation, not as the final contribution.","feed_headline":"Simulator test: Neuralangelo rebuilds satellite CAD shape from imagery","feed_subtitle":"A synthetic pipeline for reconstructing tumbling space targets shows a strong static-scene baseline before tackling dual motion.","key_machinery":"The key machinery is the simulation framework: Isaac Sim generates physically rendered images with ground-truth depth, pose, and lighting, while the observer spacecraft follows a 45-degree-inclined, stably bounded Clohessy–Wiltshire relative orbit whose initial conditions are derived from a rotation of the standard bounded-orbit condition. This provides camera poses that are assumed perfectly known. For the static baseline, Neuralangelo—a coarse-to-fine optimization applied to multi-hash encoding—carries the reconstruction, producing high-fidelity surface meshes that are quantitatively compared to the CAD model via signed C2M distances in CloudCompare.","core_discovery":"Using a custom simulation pipeline built on Isaac Sim and the Clohessy–Wiltshire relative-orbit equations, the authors generated synthetic monocular image sequences of three satellite CAD models (a 1RU CubeSat, DESDynI, and GOES-R) in a static scene. They then applied Neuralangelo, a neural surface reconstruction method, and compared the extracted meshes with the reference CAD models using signed cloud-to-mesh (C2M) distances. The meshes closely matched the references, with errors concentrated in small local regions and fine details preserved. The one systematic failure mode noted is that thin structures, exemplified by a GOES-R magnetometer, can be completely absent from the reconstruction.","pith_inferences":["The dynamic-scene evaluation, the paper's central advertised goal, is not yet presented; the reader should treat the static results as a proof-of-concept of the simulator and reconstruction pipeline rather than evidence about dynamic algorithms.","The assumption of perfectly known camera poses from the simulator is the main barrier to transferability: in a real uncooperative RPO mission, pose estimation error would likely be the dominant source of reconstruction error, so a natural test is to add pose noise or run structure-from-motion on the same imagery.","If the dynamic evaluation proceeds with the same metric, the missing-thin-structures failure mode may be even more pronounced for a tumbling target, where fast motion and occlusion can cause appendages to be imaged only briefly.","One practical extension would be using the reconstructed static mesh as an initialization prior for dynamic Gaussian splatting, similar to existing coarse-shape-prior approaches, to accelerate convergence on the tumbling scenario."],"forward_implications":["If the static baseline holds, mission planners can expect that neural surface reconstruction from monocular imagery can recover satellite geometry well enough to support proximity operations, provided thin appendages are not mission-critical.","The 45-degree inclined relative orbit ensures that top and bottom surfaces of the target are imaged, which is a necessary condition for complete reconstruction; the paper suggests this trajectory design generalizes to other targets.","The evaluation pipeline, with its ground-truth poses and CAD references, provides a controlled way to compare dynamic reconstruction algorithms (D-NeRF, Nerfies, TiNeuVox, HexPlane) on a tumbling target, which is the paper's advertised next step.","The current C2M metric fails to penalize missing geometry, so the authors propose reversing the distance direction (reference-to-reconstruction) to catch such omissions; this would make the error metric more reliable for RSO characterization.","Neuralangelo's approximately 8-hour runtime per reconstruction indicates that the static baseline is not real-time, but the paper does not claim operational speed, only fidelity."],"supporting_citations":[{"why":"Provides Neuralangelo, the reconstruction algorithm that produces the high-fidelity meshes used in the static baseline.","marker":"[37]"},{"why":"Supplies the classical orbit-element-difference condition for stably bounded relative orbits, used to set the observer's initial conditions.","marker":"[36]"},{"why":"Defines the Hill frame and the Clohessy–Wiltshire relative motion model used to simulate the observer fly-around.","marker":"[35]"},{"why":"Source of some of the NASA 3D CAD models (e.g., CubeSat, DESDynI) that serve as ground-truth geometry.","marker":"[32]"},{"why":"Alternate NASA 3D model repository providing additional CAD models used in the simulation.","marker":"[33]"},{"why":"TurboSquid supplies one of the CAD models (GOES-R) used for evaluation.","marker":"[34]"},{"why":"CloudCompare provides the cloud-to-mesh distance computation used as the geometric accuracy metric.","marker":"[39]"},{"why":"Prior work on high-fidelity 3D reconstruction of space bodies using NeRF, which this research explicitly builds upon by relaxing the static-scene constraint.","marker":"[17]"}],"fun_headline_variants":["Neuralangelo rebuilds satellite CAD from synthetic views","Synthetic satellite test: 3D mesh close to CAD, with gaps","Thin satellite parts vanish in Neuralangelo reconstruction","Isaac Sim pipeline assesses satellite 3D reconstruction","Static satellite scenes: Neuralangelo shows reconstruction promise"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The pipeline assumes that the camera poses are known exactly from the simulator's ground truth, so the reconstruction never has to contend with pose-estimation error; in a real uncooperative rendezvous mission, pose uncertainty is a primary error source and likely to dominate reconstruction quality.","fun_headline_variants_meta":{"raw":{"variants":["Neuralangelo rebuilds satellite CAD from synthetic views","Synthetic satellite test: 3D mesh close to CAD, with gaps","Thin satellite parts vanish in Neuralangelo reconstruction","Isaac Sim pipeline assesses satellite 3D reconstruction","Static satellite scenes: Neuralangelo shows reconstruction promise"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000189,"raw_usage":{"total_tokens":1149,"prompt_tokens":699,"completion_tokens":450,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":443,"completion_tokens_details":{"reasoning_tokens":371}},"tokens_in":443,"tokens_out":450,"duration_ms":6035,"temperature":1.0,"reasoning_tokens":371,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T21:27:55.923889+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same Neuralangelo reconstruction on real imagery of an uncooperative spacecraft (e.g., from a ground-based telescope or an inspection mission) where camera poses must be estimated from the images themselves; if the mesh-to-CAD error increases sharply or thin structures disappear entirely, the simulator-only transferability claim is falsified. A simpler check within the paper's own framework: add realistic pose noise to the ground-truth camera positions and measure how C2M error degrades.","supporting_citations":[{"cited_title":"Neuralangelo: High-fidelityneuralsurface reconstruction,","cited_arxiv_id":null,"evidence_quote":"Provides Neuralangelo, the reconstruction algorithm that produces the high-fidelity meshes used in the static baseline."},{"cited_title":"Relative orbit geometry through classical orbit element differences,","cited_arxiv_id":null,"evidence_quote":"Supplies the classical orbit-element-difference condition for stably bounded relative orbits, used to set the observer's initial conditions."},{"cited_title":"L.,Analytical Mechanics of Space Systems, American Institute of Aeronautics and Astronautics, Incorporated, 2014","cited_arxiv_id":null,"evidence_quote":"Defines the Hill frame and the Clohessy–Wiltshire relative motion model used to simulate the observer fly-around."},{"cited_title":"Accessed: May 08, 2025","cited_arxiv_id":null,"evidence_quote":"Source of some of the NASA 3D CAD models (e.g., CubeSat, DESDynI) that serve as ground-truth geometry."},{"cited_title":"Accessed: May 08, 2025","cited_arxiv_id":null,"evidence_quote":"Alternate NASA 3D model repository providing additional CAD models used in the simulation."},{"cited_title":"Accessed: May 20, 2023","cited_arxiv_id":null,"evidence_quote":"TurboSquid supplies one of the CAD models (GOES-R) used for evaluation."},{"cited_title":"CloudCompare,","cited_arxiv_id":null,"evidence_quote":"CloudCompare provides the cloud-to-mesh distance computation used as the geometric accuracy metric."},{"cited_title":"High-Fidelity3DReconstructionofSpaceBodiesUsingMachineLearningandNeuralRadianceFields,","cited_arxiv_id":null,"evidence_quote":"Prior work on high-fidelity 3D reconstruction of space bodies using NeRF, which this research explicitly builds upon by relaxing the static-scene constraint."}],"review_version":1}