acceptodds
Under review as a conference paper at ICLR 2027

DynaBR: Dynamics-Integrated Base–Residual Learning Framework for Non-Cooperative Satellite Orbit Forecasting

Abstract

Growing congestion in low Earth orbit (LEO) makes accurate orbit forecasting essential for spacecraft safety. For non-cooperative targets, forecasters rely solely on external observations. Uncertain target properties and environments limit physics-based perturbation modeling, while data-driven models fit complex patterns but struggle to incorporate known dynamics. We first construct a dataset covering 2,000 LEO targets, including satellites, rocket bodies, and space debris, by propagating discrete two-line element (TLE) records into regularly sampled orbital state sequences. We then propose DynaBR, a unified forecasting model. Global Dynamics Reference provides the dominant orbital structure; Local Dynamics Alignment calibrates future references using historical phase discrepancies; and Dynamics-Conditioned Residual Forecasting predicts corrections from historical residuals and aligned future references to compensate for perturbation effects that approximate dynamics cannot explain. Experiments show that DynaBR achieves the strongest overall forecasting performance on unseen targets, achieves the strongest overall performance across forecasting horizons while substantially outperforming both physics-based and data-driven methods at longer horizons and across all target types. Over a 48-hour horizon, DynaBR achieves an average relative reduction of 99.45% in mean position error across 12 SOTA baselines.

Primary area:
learning on time series and dynamical systems
Keywords:
Orbit Prediction ;Physics-Informed

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_i1VwOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper targets long‑horizon LEO orbit forecasting for non‑cooperative space objects such as satellites, rocket bodies and debris. It proposes DynaBR, a hybrid physics‑data framework composed of three modules: Global Dynamics Reference, Local Dynamics Alignment (LDA), and Dynamics‑Conditioned Residual Forecasting. The core contribution is LDA, which calibrates phase drift via linear fitting of historical phase discrepancies in the RTN frame and reduces long‑horizon error substantially.

The authors build a large dataset from real TLE archives by propagating discrete TLEs into regularly sampled time series using SGP4. Train/val/test splits are object‑disjoint to evaluate generalization to unseen targets. Experiments compare DynaBR against pure physics baselines and many time‑series deep learning models across forecast horizons up to 48 hours. Ablations verify the effectiveness of each component, especially LDA. Case studies on ISS conjunction assessment demonstrate downstream application potential.

Critical strengths
  1. It addresses a critical pain point in orbit forecasting: approximate dynamical models produce time‑accumulated orbital phase drift, which causes error explosion in long‑horizon forecasting. Instead of directly estimating hard‑to‑obtain physical parameters (area‑to‑mass ratio, attitude), it fits the linear trend of phase discrepancies from historical orbital time‑series, and applies rotational correction to future reference orbits within the local RTN orbital frame.

  2. The paper builds a large‑scale LEO dataset covering satellites, rocket bodies and space‑debris with object‑disjoint evaluation.

Critical weaknesses
  1. While the paper states that it targets non‑cooperative space objects, all validation experiments are performed on data simulated and generated by the SGP4 propagator. It appears that all train/val/test trajectories are reconstructed by propagating discrete TLEs with SGP4. Thus, both inputs and ground‑truth labels are SGP4 outputs instead of real‑world physical orbits. The model may fit SGP4‑specific artifacts and biases rather than true unmodeled orbital perturbations. No held‑out independent precise ephemeris is used to test out‑of‑SGP4 generalization, leading to considerable simulation‑to‑reality risk. Additionally, the dataset produces ideal dense uniform time‑series and ignores noise from real‑world orbit‑determination pipelines; reported performance may drop under real operational inputs.

  2. I cannot see the LEO dataset released.

Questions for authors
  1. It is clear that the dataset is filled‑in from TLE records via the SGP4 propagator. Why is the latter barely mentioned in the main text?

  2. Does the manuscript suffer from substantial instances of the following writing‑style issues, and how will the authors fix them? These include missing full‑name expansion for abbreviations at their first appearance in the main text, and inconsistent capitalization of terms that disregards standard academic conventions.

  3. SGP4 is merely a simplified analytical physical propagator. Many real-world effects, including time-varying perturbations, maneuvers, drag, and unanticipated orbital changes, cannot be precisely modeled by SGP4. Does generating orbits using SGP4 oversimplify the problem? Is this analogous to stock trajectories: phenomena that are inherently hard to predict yet simplified for forecasting?

  4. The LDA assumes linear phase discrepancy within the fitting window. How robust is this linear assumption for debris with rapidly changing drag? Could non-linear phase bias degrade LDA performance significantly?

6 comments

Thank you for your comments and we have conducted additional experiments. Our responses are provided below.

Q1: Why is the role of SGP4 under-emphasized in the main text?

Thank you for noting this. Because TLE data are conventionally processed with the Simplified Perturbations models (SGP4), we treated the propagation step as background knowledge and did not highlight it sufficiently. This was an oversight, we will make the data-generation pipeline explicit in the revised manuscript. The full dataset will be released upon acceptance.

Q2: Missing abbreviation expansions and inconsistent capitalization.

We apologize for these writing-style issues. In the revision we will expand LEO, TLE, SGP4, TEME, RTN and LDA at first use and enforce consistent capitalization of module names throughout.

Q3: Does generating trajectories with SGP4 oversimplify the problem?

We appreciate this important concern. SGP4 is a medium-fidelity analytical propagator that accounts for the dominant LEO perturbations (non-spherical gravity, atmospheric drag, third-body effects, and resonance terms) while omitting solar radiation pressure, tides, and higher-order gravity. For the majority of LEO objects, the neglected terms are small relative to drag on the time scales we consider; published SGP4 residual errors are typically on the order of 1 km.

Publicly available continuous trajectories for non-cooperative objects are almost exclusively obtained by propagating sparse observations (TLE or VCM) with a similar dynamical model. High-precision GNSS-derived ephemerides are generally available only for cooperative assets. Consequently, the SGP4-derived dataset remains the most practical large-scale public alternative for the non-cooperative setting we target.

To further mitigate the simulation-to-reality concern, we additionally validated against independent Sentinel POD products. These Precise Orbit Determination solutions are derived from on-board GNSS observations and achieve typical 3D RMS accuracies of a few centimeters (POEORB products typically ~0.6 cm), providing a high-accuracy external reference. The mean position errors of the SGP4 on these POD trajectories are shown below:

Forecast Horizon1 h6 h12 h24 h48 h
SGP4 Mean Position Error (km)0.7190.8710.8901.0211.520

Q4: Robustness of the linear phase-discrepancy assumption under rapidly varying drag.

Thank you for highlighting this assumption. LDA uses a phase-drift rate estimated from historical states to locally calibrate the reference trajectory. Our experiments show that this simple correction is effective on the original validation set and can retain substantial benefits even when the phase discrepancy contains a nonlinear component.

On the original 200-object, 1,000-window validation set, linear LDA achieves a mean position error of 2.562 km, compared with 4.334 km for fixed quadratic extrapolation. The quadratic variant increases the error by 1.772 km, with an object-paired bootstrap 95% CI of [0.285, 4.424] km. This result supports the low-order calibration used in our framework. Building on the linear reference, DynaBR further reduces the mean error to 1.759 km, a 31.3% reduction.

We agree that rapidly varying drag can challenge linear phase extrapolation. When the observed history contains phase curvature that persists into the forecast interval, higher-order extrapolation or a learned residual correction may improve predictions. However, if drag changes after the last observation in ways that cannot be inferred from the available history, history-only methods share an information constraint, although their errors may differ. Additional observations can support forecast updates, while external density forecasts or information about the object’s drag response may improve prediction. Our current results do not establish robustness to rapidly varying physical drag.

To examine sensitivity to phase curvature directly, we prescribed

ϕ(t)=A(u+cu2),u=t/(48 h),A=0.001 rad\phi(t)=A(u+cu^2),\qquad u=t/(48\,\mathrm{h}),\quad A=0.001\,\mathrm{rad}

where c denotes the ratio of the quadratic to linear contribution at 48 h. We retained the original LDA procedure: 2,048 historical samples, a 1,280-point fitting window, one-minute sampling, and a 48 h forecast. The same phase law applies throughout history and forecast, and fitting uses historical samples only.

Quadratic-to-linear ratio ccUncorrected phase MAE (mrad)LDA phase MAE (mrad)Error reduction
00.500≈0100.0%
0.250.5840.13976.2%
0.50.6670.27858.3%
10.8340.55633.4%
21.1671.1114.8%

LDA reduces the mean phase error by 76.2% and 58.3% at c=0.25c=0.25 and c=0.5c=0.5, respectively. Even when the quadratic contribution equals the linear contribution at 48 h (c=1c=1), it retains a 33.4% reduction.

AuthorsOpenReview ↗

Follow-up on Our Rebuttal to Reviewer

Dear Reviewer i1Vw,

I hope this message finds you well. As the discussion period is nearing its end with less than one day remaining, I wanted to ensure we have addressed all your concerns satisfactorily. If there are any additional points or feedback you'd like us to consider, please let us know. Your insights are invaluable to us, and we're eager to address any remaining issues to improve our work.

Thank you for your time and effort in reviewing our paper.

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

This paper proposes DynaBR, a hybrid framework for non-cooperative orbit forecasting that combines physics-based reference propagation, local phase alignment, and dynamics-conditioned residual learning. Experiments on a TLE-derived dataset of 2,000 objects show substantial improvements over physics-based and time-series baselines for long-horizon forecasting on unseen objects.

Critical strengths
  1. Clear integration of physical dynamics, local phase calibration, and learned residual correction.
  2. Strong results, with ablations supporting the contributions of the main components.
  3. A diverse, multi-object dataset is collected
Critical weaknesses
  1. Since the reference trajectories are reconstructed from TLEs using SGP4, the reported accuracy primarily measures agreement with these reconstructed trajectories. Validation against independent, higher-accuracy orbit, such as suitable VCM data, would strengthen the evidence for real-world forecasting accuracy.

  2. Why does TimesFM-3 perform poorly on this data? Please give more insights like whether this reflects insufficient input context, the long-horizon decoding strategy, etc.

  3. Does the dataset include documented orbital maneuvers or substantial changes in orbital drift? Evaluating these cases separately, particularly post-maneuver adaptation, would better establish the method’s robustness and practical applicability.

  4. How were the input window and forecast horizon steps selected? Please explain their relationship to LEO orbital periods, lunar periodicity, or other relevant perturbation timescales, and assess sensitivity to these choices.

  5. Please report training cost, memory usage, and end-to-end inference latency.

  6. The evaluation metrics mainly focus on position and velocity errors. It does not directly assess the physical consistency of the final predicted trajectories.

Questions for authors

see the weaknesses section

2 comments

Thank you for your comments and we have conducted additional experiments. Our responses are provided below.

W1: Validation against independent high-accuracy orbits (e.g., VCM or POD) is needed.

Thank you for this suggestion. VCM products are themselves discrete observations that must be propagated; we therefore turned to independent POD products from the ESA’s Sentinel missions, which are derived from on-board GNSS observations.

We kept the trained DynaBR model completely frozen and evaluated on these POD trajectories. For the SGP4 baseline we re-fitted a TLE to the same historical window.

MethodMean position error (km)
SGP41.520
Ours0.828

Average first over windows of each satellite, then equal-weight average of the two satellites. DynaBR reduces mean position error by approximately 45.5 % relative to SGP4, indicating that the advantage persists under an independent, higher-accuracy reference.

W2: Why does TimesFM-3 perform poorly?

We re-examined the TimesFM-3 configuration (full 2 048-point history, six-variable input, median forecast) and analyzed error growth by horizon:

Horizon0–6 h6–12 h12–18 h18–24 h24–30 h30–36 h36–42 h42–48 h
Pos. mean (km)54.3580.07103.47140.95198.53259.46581.411 288.53

Error accumulates rapidly with lead time, consistent with the difficulties of long-horizon autoregressive decoding on this highly structured orbital dynamics.

W3: Does the dataset contain documented maneuvers, and how does the method behave post-maneuver?

Using ESA’s official maneuver logs (burn start/stop times), we performed a maneuver sensitivity study on Sentinel. The model remained frozen. Forecasts were initiated (i) 6 h before the first burn and (ii) 6 h after the last burn-out; both methods received only the history available at the respective origin, and SGP4 was re-fitted to that history. Errors were computed only on non-thrust, quality-flagged POD states.

Forecast originOurs (km)SGP4 (km)
6 h before first burn2.4352.650
6 h after last burn-out4.9845.784

Both methods are affected by maneuvers; DynaBR retains advantage in both the during-maneuver and post-maneuver regimes.

W4: Rationale for the chosen history length and forecast horizon.

The 2048-point history (≈ 34.1 h) spans multiple LEO orbital periods. For example, a typical 500 km sun-synchronous orbit has a period of approximately 94.6 min (≈ 1.58 h), so the history covers roughly 21–22 revolutions. This provides sufficient context for estimating recent phase drift and recurring residual patterns while remaining compatible with the length-64 block encoder. The 2880-point (48 h) forecast horizon corresponds to about 30 revolutions and is long enough to expose error accumulation of the pure dynamics reference and to test whether the learned residual correction remains effective across many orbital periods.

W5: Training cost, memory, and inference latency.

DynaBR contains 350982 trainable parameters. All residual models were trained for 50 epochs (80 000 windows, 5 000 updates) on a shared RTX 5090.

ModelTraining time (s)Peak training GPU memoryInference latency (ms)
DLinear265.21–265.41542.3 MiB862.55
TimesNet4823–56261588–1592 MiB847.93
Ours265.39–265.57414.6 MiB703.85

W6: Physical consistency of the predicted trajectories.

We evaluated kinematic consistency between the predicted position sequence and the predicted velocity by comparing the numerical time derivative of position with the model’s velocity output. Inconsistency is defined as Dtrk=rk−2−8rk−1+8rk+1−rk+212Δt,Δt=60 s.D_t\mathbf r_k = \frac{\mathbf r_{k-2}-8\mathbf r_{k-1}+8\mathbf r_{k+1}-\mathbf r_{k+2}} {12\Delta t}, \qquad \Delta t=60\ \mathrm{s}. End-points are discarded and only samples that do not cross a TLE-source discontinuity are scored.

MethodMean inconsistency (m/s)p95 (m/s)
Ours0.434 ± 0.0071.696 ± 0.093
DLinear0.878 ± 0.1433.329 ± 0.694
TimesNet0.753 ± 0.0402.928 ± 0.194

DynaBR reduces mean inconsistency by 50.6 % and 42.3 % relative to DLinear and TimesNet, respectively, showing that the accuracy gains are accompanied by improved position–velocity compatibility.

Thanks for providing more information and results. I'd like to keep my positive evaluation.

Reviewer_DFhUOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper forecasts non-cooperative LEO objects 48 hours ahead from state histories alone. It constructs a 2,000-object dataset by propagating catalog TLEs with SGP4. DynaBR combines a WGS72 J2 reference, history-fitted phase alignment and Transformer residual correction. On 200 test objects, it reports a mean position error of 0.975 km against 17.662 km for J2 propagation, with ablations and an ISS conjunction case study. Limitation. TLE-derived SGP4 trajectories may be the best public reference for non-cooperative objects, but they remain a catalog-derived proxy. Error against them establishes agreement with catalog propagation rather than physical orbit accuracy.

Critical strengths

Local Dynamics Alignment is a simple and interpretable mechanism for adapting a shared physics reference to each object. It estimates recent phase drift directly from the observed history and adjusts the future reference without requiring unavailable target-specific physical parameters. The reversed-sign control helps show that the fitted direction of the phase correction is meaningful. This provides direct support for Local Dynamics Alignment within DynaBR.

Critical weaknesses

W1. Equation 14 generates each target state with the latest TLE available at that minute. The 48-hour target begins with SGP4 propagation from the TLE available at the forecast origin, then switches to later TLEs that incorporate new observations unavailable to DynaBR. It is therefore a stitched catalog trajectory rather than one open-loop SGP4 rollout. The reported 0.975 km error mixes agreement with open-loop propagation before the first update and prediction of catalog corrections over all subsequent intervals. Please add an SGP4 baseline that fits mean elements and B$^*$ to the same history states given to DynaBR, then forecasts without later TLEs. Report errors over the interval before any future TLE is incorporated, and separately over all subsequent intervals whose targets use future TLE updates. This distinguishes open-loop propagation from prediction of catalog corrections.

W2. Table 1 compares methods with unequal inductive biases. Most learned baselines predict the full 48-hour trajectory directly from data, whereas DynaBR starts from PertDyn, aligns that physics-derived reference to the observed history and learns only its residual correction. Its advantage over purely data-driven methods may therefore arise from the physics prior rather than DynaBR's specific architecture. No comparison establishes whether its alignment and residual-learning design improves over a method with a comparable physics-based inductive bias. The paper cites relevant hybrid approaches by Curzi et al., AutoOrbit and Yin et al. Please compare DynaBR against competitive hybrid methods that receive the same history and a comparable physics-based inductive bias, using multiple seeds and comparable tuning budgets.

W3. Before Section 4.4, DynaBR is defined and evaluated as a deterministic point forecaster. Equations 11 and 24 optimize position and velocity errors, the model produces no predictive distribution or object-specific covariance, and Table 1 reports only point-error summaries. Section 4.4 then mentions post hoc covariance forecast uncertainty. The covariance pools validation residuals across 200 objects and five windows and therefore describes population-level error dispersion for each method, not uncertainty predicted for the ISS encounter. The nominal 95% ellipsoid further assumes an isotropic Gaussian distribution, yet no held-out coverage or covariance-realism test is reported. Such tests are central to probabilistic forecast evaluation [1] and probability-based conjunction analysis [2]. The resulting collision probability is therefore a sensitivity calculation using aggregate residual statistics, not a calibrated object-specific risk estimate. Please present Section 4.4 as an error-dispersion and collision-sensitivity analysis. If predictive uncertainty is claimed, output an object-conditional distribution or covariance and evaluate it with a strictly proper scoring rule. Please report calibration and sharpness of object-specific time-varying ellipsoids on held-out datasets from the future against competitive baselines.

[1] T. Gneiting and A. E. Raftery. “Strictly Proper Scoring Rules, Prediction, and Estimation.” Journal of the American Statistical Association, vol. 102, no. 477, pp. 359–378, 2007. https://doi.org/10.1198/016214506000001437

[2] NASA. Spacecraft Conjunction Assessment and Collision Avoidance Best Practices Handbook, Appendix I, “Satellite Covariance Realism Assessment Procedures.” NASA CARA Handbook

Questions for authors

See Weaknesses.

2 comments

Thank you for your comments and we have conducted additional experiments. Our responses are provided below.

W1: The ground-truth trajectory stitches later TLEs, an open-loop SGP4 baseline fitted only to the same history is needed, together with pre-/post-update evaluation.

Thank you for this precise methodological suggestion. We agree that the original evaluation mixed open-loop propagation with prediction of catalog updates.

We therefore added a History-fit SGP4 baseline that receives exactly the same position–velocity–time history given to DynaBR, fits mean elements and B* from the terminal state, and then performs a pure 48 h open-loop propagation. No future TLEs are ever supplied to either method. We report four complementary metrics: the full forecast interval, the interval before the first future TLE update, the interval from the first update to the end.

Position mean (km)Full intervalBefore first updateAfter first update
Ours1.7590.6342.316
SGP41.5000.4502.394

SGP4 retains a clear advantage on windows without any TLE update. Once a catalog update occurs inside the forecast horizon, DynaBR achieves lower error, indicating that the residual learner captures information beyond pure open-loop propagation of the fitted TLE.

W2: Advantage may stem purely from the physics prior; comparison against published hybrid methods under matched inductive bias is required.

We agree that a fair comparison should hold the physics-based inductive bias constant. We carefully examined the three hybrid methods you mentioned with respect to input constraints and training methods.

  • Curzi et al. train a separate residual network per target. Our evaluation method requires a single shared model that generalizes to completely unseen objects; introducing per-target fine-tuning would change the experimental setting.
  • AutoOrbit constructs its periodic reference from ground-track repeat cycles whose historical span exceeds our 34.1 h input window, and its physics loss requires mass and frontal area—quantities unavailable under the six-dimensional state-only input of the present task. Shortening the reference period or assigning a uniform area-to-mass ratio would alter the method’s core mechanisms; we therefore could not produce a faithful comparison under the shared input constraint.
  • Among the hybrid methods mentioned, we found that APHYNITY (Yin et al.) can be fairly compared under the same limited input setting. We adapted it by using the identical force model as the physics term, predicting an acceleration residual from a history-conditioned network, and integrating continuously for 48 h from the common forecast origin without observation resets. Both methods shared the same training windows, 50-epoch schedule, 5 000 optimization steps, validation method, and three random seeds.
MethodPosition mean (km)Velocity mean (m/s)
APHYNITY15.662 ± 1.13515.518 ± 1.158
DynaBR1.765 ± 0.0051.633 ± 0.011

The large gap under matched physics and training budget further supports the contribution of phase alignment, residual architecture, and future-dynamics conditioning inside DynaBR.

W3: Population-level residual covariance cannot be interpreted as object-specific predictive uncertainty or calibrated collision risk.

Thank you for highlighting this distinction. We agree and would like to clarify the intended claim of Section 4.4. DynaBR is a deterministic point forecaster, and this analysis characterizes the population-level dispersion of its forecast errors and the implications for a controlled conjunction calculation.

Section 4.4 explicitly estimates object-balanced covariances from validation residuals and holds the nominal encounter geometry fixed across methods. Its principal finding is that DynaBR reduces empirical spatial error dispersion, yielding more concentrated error statistics on the evaluated cohort. This scope is already reflected in Appendix A.6, "Forecast-Error Dispersion and Conjunction Sensitivity," and A.6.2, "Controlled Collision-Probability Sensitivity." The probability calculation illustrates how aggregate error covariance affects the probability integral under a common geometry.

Thanks for addressing my concerns. I would raise my score accordingly.

Reviewer_Nnz9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes DynaBR, a physics–learning framework for forecasting the orbits of non-cooperative LEO objects. It constructs a TLE-derived dataset of 2,000 objects by using SGP4 to generate regularly sampled position and velocity sequences. Given approximately 34 hours of historical states, the model predicts the next 48 hours. DynaBR generates a physics-based reference trajectory, adjusts its future phase using historical phase discrepancies, and applies a Transformer-based residual predictor conditioned on the aligned future reference and forecast lead time. Experiments use object-disjoint training, validation, and test sets and compare against physics-based, data-driven, and pretrained time-series models. The reported results show lower position and velocity errors over the 48-hour forecast interval, with ablation studies examining the contributions of phase alignment and residual forecasting.

Critical strengths

The paper addresses a relevant problem in forecasting the orbits of non-cooperative objects. Using a physics-based reference and learning corrections is a reasonable approach, and the phase alignment step provides a way to adapt the reference to each object's recent history without requiring detailed physical properties.

The dataset covers 2,000 LEO objects, with separate objects for training, validation, and testing. This allows evaluation on unseen objects rather than only future segments of the same objects. The reported results show improvements over the included baselines across different target types. The ablation studies also help distinguish the effects of phase alignment and neural residual correction.

Critical weaknesses

(1) The ground truth is generated from TLE records using SGP4, but the physics-based baselines only include two-body and J2 models. I think an important comparison is missing here: taking the latest TLE available at the forecast cutoff and propagating it with SGP4 for the next 48 hours. Without this baseline, it is hard to tell how much the method improves over standard TLE-based forecasting, rather than just correcting a simplified physics model.

(2) The modeling contribution seems limited, as the method mainly combines a physics-based reference, phase alignment, and a Transformer for residual forecasting. Although other AI models are compared as direct forecasting baselines, only the Transformer is tested as the residual predictor within this framework. It is unclear whether replacing it with another model would give similar results, so the benefit of this particular modeling choice is not well established.

Questions for authors

(1) How does DynaBR compare with propagating the latest TLE available at the forecast cutoff using SGP4 for the next 48 hours? This comparison would help clarify the improvement over standard TLE-based forecasting.

(2) Have you tested other residual predictors, ranging from simple models to strong state-of-the-art forecasting models, using the same physics-based reference and phase alignment? This would help clarify whether the proposed Transformer-based predictor offers a specific advantage or whether other models would achieve similar results.

3 comments

Thank you for your comments and we have conducted additional experiments. Our responses are provided below.

Q1 / W1: Comparison with open-loop SGP4 propagation from the latest TLE available at the forecast origin.

Thank you for highlighting this important baseline. We agree that comparing against standard TLE-based forecasting is essential for contextualizing our results.

We have added an open-loop SGP4 baseline that uses only the latest TLE whose epoch is no later than the forecast origin and propagates it for 48 h without receiving any subsequent TLEs. Because our ground-truth targets equal this open-loop SGP4 trajectory whenever no TLE update occurs inside the forecast window, we focus evaluation on windows that do contain at least one TLE update and report errors after the first update.

MethodMean position error after TLE update (km)
SGP40.646
Ours0.638

DynaBR’s mean position error is slightly lower than Latest-TLE SGP4. The two methods therefore exhibit comparable accuracy on this strict open-loop setting.

Q2 / W2: Would other residual predictors yield similar results under the same physics reference and phase alignment?

Thank you for raising this ablation question. To isolate the contribution of our residual predictor, we replaced the Transformer with DLinear and TimesNet while keeping the identical physics reference, Local Dynamics Alignment (LDA), residual target, future-reference query, output reconstruction, training schedule, and model-selection protocol. Each model was trained for 50 epochs (80 000 windows, 5 000 optimization steps) with three random seeds.

MethodPosition mean (km)Position p95 (km)Velocity mean (m/s)
DLinear2.331 ± 0.0086.520 ± 0.0171.951 ± 0.012
TimesNet2.328 ± 0.0076.537 ± 0.0181.930 ± 0.013
Ours1.765 ± 0.0052.523 ± 0.0391.633 ± 0.011

Relative to DLinear and TimesNet, DynaBR reduces mean position error by 24.3 % and 24.2 %, and mean velocity error by 16.3 % and 15.4 %, respectively. These results indicate that the modeling choice of the residual predictor still contributes under a shared physics-and-alignment framework.

To further test the value of conditioning on the future reference state, we added a Transformer variant that receives only the forecast lead time as query. All other components remain identical.

HorizonTransformer p95 (km)Ours p95 (km)
0–6 h1.2801.354
6–12 h1.5981.579
12–24 h2.3142.022
24–48 h4.2493.267

The benefit of dynamics-conditioned residual prediction becomes more pronounced at longer horizons, supporting the design of our residual module.

Thank you for providing both additional experiments. They have addressed my concern about using other models instead of the Transformer. However, the improvement over SGP4 is quite small, and it is still unclear wheter this small difference is statistically significant. DynaBR has a relatively complex structure while SGP is a simple physical model. So I still think it is difficult to say that DynaBR is much better model that SGP4.

AuthorsOpenReview ↗

Further Comparison of Our Method DynaBR with SGP4.

Thank you for your follow-up and valuable comments. To further evaluate the practical applicability of DynaBR to long-term orbit forecasting of non-cooperative space objects, we conducted an additional long-horizon comparison using precise Sentinel POD products as the ground-truth reference.
To simulate a realistic prediction scenario in which future observations of non-cooperative objects are unavailable, we evaluated DynaBR through successive 48-hour autoregressive rollout blocks, feeding its own predictions back as the input history for subsequent blocks. No future observations or external state corrections are used throughout the rollout. The SGP4 baseline follows the same forecasting setting as former experiment. At each forecast horizon, we compute the RMS of the three-dimensional position errors across windows, with equal weighting within each spacecraft and across spacecraft.

Forecast horizonDynaBR RMS (km)SGP4 RMS (km)RMS reduction
2 days2.0861023.65627542.94%
4 days5.6011498.81389436.45%
6 days9.83909318.27886346.17%
8 days19.01247933.18991242.72%
10 days34.97671050.31805430.49%

DynaBR consistently achieves lower position RMS than SGP4 across all five evaluated horizons, with reductions ranging from 30.49% to 46.17%. Notably, its advantage persists even after ten days of fully autonomous autoregressive forecasting, reducing the position RMS from 50.318 km to 34.977 km.
These results provide additional evidence of DynaBR’s robustness in long-term prediction without observation updates. More importantly, while SGP4 exhibits increasingly pronounced error accumulation and orbital divergence over extended propagation periods, DynaBR maintains substantially lower error growth.
We attribute this advantage to the effective fusion of physical orbital dynamics and data-driven machine learning, which combines the structural priors of orbital dynamics with the adaptive modeling capability of neural networks. This synergy enables DynaBR to better capture complex orbital evolution patterns, mitigate accumulated prediction errors during autoregressive rollout, and substantially slow long-term error growth compared with conventional analytical propagation in observation-limited forecasting scenarios.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.