acceptodds
Withdrawn from ICLR 2027

From Historical Errors to Future Trajectories: Learning Transferable Knowledge for Satellite Orbit Prediction

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Satellite orbit prediction is fundamental to modern space operations, with an increasing demand for accurate trajectory forecasts driven by the rapid expansion of satellite constellations. Recent studies employ deep learning to model the residuals between predicted and true trajectories, providing a promising solution for orbit forecasting. However, such residual learning paradigms simply treat prediction errors as supervision, overlooking the inherent periodicity of orbital motion and thereby limiting their predictive performance. To this end, we introduce OrbitEcho, a novel residual correction framework that encodes historical prediction experience into transferable knowledge for accurate trajectory forecasting. Specifically, OrbitEcho first stores forecasting errors from preceding orbital cycles in a memory bank. It then retrieves phase-aligned trajectory errors at similar orbital states as informative references for residual correction. For each retrieved candidate, OrbitEcho further exploits orbital periodicity to model its residual evolution over one orbital cycle, which enables historical errors to be adapted to the current prediction. Together, OrbitEcho distills historical predictions into two complementary forms of transferable knowledge, i.e., trajectory errors and residual dynamics, which are jointly leveraged to refine orbit forecasts. Extensive experiments on real-world satellites demonstrate the superior performance of OrbitEcho, reducing the average MAE from 8.40 km for a state-of-the-art learning-based baseline to 0.31 km, yielding an order-of-magnitude improvement. OrbitEcho also exhibits strong generalization to unseen satellites and robust performance under severe geomagnetic storms.

Primary area:
learning on time series and dynamical systems
Keywords:
Satellite orbit prediction, Time-series Forecasting, Residual Correction, Orbital Dynamics

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_CWfUOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents ORBITECHO, a residual‑correction framework for LEO satellite orbit prediction. Built upon a base time‑series predictor that outputs reference trajectories, the method maintains a memory bank storing historical prediction residuals (errors between reference forecasts and ground‑truth orbits). It retrieves phase‑aligned historical residual cases leveraging orbital periodicity, then models cross‑cycle residual dynamics to adapt retrieved errors for current correction; a lightweight query‑key router performs adaptive aggregation. Theoretical analysis provides smoothness‑based justification for residual transferability.

Critical strengths

The paper designs a phase‑aligned memory‑retrieval pipeline that reuses historical prediction residuals across orbital cycles, together with cross‑cycle residual‑dynamics adaptation.

Critical weaknesses
  1. The largest performance jump comes from phase‑aligned residual matching, a hand‑engineered prior. The neural‑network components are minimal: the base predictor is off‑the‑shelf time‑series backbone, and the query‑key router is only a lightweight aggregation module. Therefore the possible contribution for ICLR is relatively limited.

  2. The claimed zero‑shot generalization and robustness are demonstrated exclusively on cooperative satellites with access to IDS precise ephemerides. Zero‑shot here means no network fine‑tuning, yet each new satellite still requires ground‑truth trajectories to compute residuals and populate the memory bank. No evidence is provided for scenarios without precise ground truth, such as non‑cooperative debris. Hence the generalization and robustness cannot be directly extrapolated to general SSA missions.

Questions for authors
  1. Could you pinpoint your single most central technical contribution that enables this remarkable result: reducing the average MAE from 8.40 km for a state‑of‑the‑art learning‑based baseline down to 0.31 km?
  2. Since you characterize historical prediction experience as transferable knowledge, where exactly does this "transferable" property manifest? This appears to be merely an empirical pattern, just the residual difference between predicted reference trajectories and ground‑truth true trajectories? Why is the term "transferable" adopted here?
  3. The residual knowledge stored in the memory bank relies on access to true satellite trajectories from IDS post‑processed precise ephemerides. Since such ground‑truth trajectories are unavailable for non‑cooperative space objects, what are the practical boundaries of the proposed method? Can it be adapted to work with only orbit‑determination estimates instead of true trajectories?
  4. Could you quantify how much gain originates from phase‑aligned retrieval versus the learnable query‑key residual‑dynamics router?
Reviewer_tYunOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes ORBITECHO, which refines satellite orbit predictions by retrieving phase-aligned historical errors from the preceding three orbital cycles and learning to aggregate their cross-cycle changes. Using precise orbit products from the International DORIS Service (IDS), the authors train and evaluate the model on three satellites and assess generalization on two unseen satellites. The method achieves higher accuracy than the evaluated general time-series and orbit-specific learning baselines.

Critical strengths
  1. The evaluation uses real-world orbit data and includes a geomagnetic-storm case study.

  2. Comparisons cover diverse baselines.

  3. The paper is clearly written and easy to follow.

Critical weaknesses
  1. Using recent historical prediction errors as inputs may pose practical challenges.

  2. The presentation and supporting results regarding transferability could be improved.

Questions for authors
  1. The method requires prediction errors from the most recent three orbital cycles, but their availability in real time is unclear. How are sufficiently accurate reference states obtained to compute these errors, and does the evaluation account for delays in their availability?

  2. The distinction between temporal error reuse and cross-satellite knowledge transfer is unclear. What exactly is transferred to an unseen satellite, and does its memory bank contain that satellite’s own historical errors? Which reported results distinguish the contribution of the learned rule from simply reusing target-specific history?

  3. The protocol for forecasts extending beyond the input window is unclear. Are these forecasts generated in one shot or through rolling updates with new observations, and is phase alignment recomputed?

  4. The high-fidelity numerical propagator could achieve lower errors, but it is absent from the efficiency comparison.

  5. The model corrects both position and velocity, but the evaluation reports only position errors.

Reviewer_G7dhOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents an approach to orbit prediction, following previous work in the residual correction paradigm, it presents a technique that finds the “closest” error correction (considering the orbit period) and uses that to improve upon a DLinear model. The correction requires identifying the ground-truth to prediction residual. It applies this model to training set satellites, but also presents zero-shot results for unseen satellites. The forecasting is evaluated for a forecasting window of 60 minutes.

Critical strengths

It evaluates both in-distribution and zero-shot orbital scenarios, with the satellites in Table 3 differing somewhat from those in the training set, further strengthening the paper's generalization claims.

It compares against many deep learning baselines.

Critical weaknesses

Below I present weakness I identified in the paper, the most relevants are the real-life application of this work, and the limited evaluation of 2,4 and 6 h, which for this use-case is still a short horizon.

NOTE: Reviewers are not required to read the appendix, and the appendix should be placed in the main, not in the supplemental materials.

(1) No dedicated related work section in the main paper, missing relevant work in retrieval augmented time-series models. ([2],etc.)

[1] Han, S.; Lee, S.; Cha, M.; Arik, S. O.; and Yoon, J. 2025. Retrieval Augmented Time Series Forecasting. In Proceedings of the 42nd International Conference on Machine Learning, volume 267 of Proceedings of Machine Learning Research, 21774–21797. PMLR.

(2) Sentinel are two twin satellites, that follow the same orbital path, with a 140º difference in phase. So the time-series are remarkably the same, between the two. Furthermore, they orbit at around 800km, which means that the orbital period is approx 100 minutes. The forecasting window does not reach one single orbital period, which hinders the applicability of this work.

(3) In line 161, the authors claim “the orbital period may vary over time”, please provide the variation of the orbital period found in the data. Tipically, for a single day, the variation is in the scale of miliseconds.

(4) In equation (7), r_{i+h} is never explictly defined.

(5) No confidence interval in results table. SGP4 uses TLE, but the authors use cartesian coordinates as the ground truth. Please evaluate the error in transforming TLE to cartesian in GCRS. to separate the error of the SGP4 propagation and the TLE data itself. Also please provide the specific link to the ground truth data and the data format (CPFv2?).

Application evaluation: As an application paper, I am required to judge this paper in is usefulness as an application. Orbital prediction is useful to predict and avoid collisions, to evaluate orbital manouvers, to schedule station-keeping manouvers or to plan future tasks ( like EO jobs).

(1) This model requires acess to a ground truth in the past, to calculate residuals. For past data this is possible, despite the very limited usefulness. For current data, meaning the goal of predicting the 60 minutes ahead of NOW, this is rarely possible, since publicly available data is published once a day. Even for a satellite owner, a ground station does not have updated state information all the time, typically only when it the passes by the gs, which is not every orbit.

(2) Finally, for planning, the most relevant and useful application, there is no available ground-truth. A proxy such as a high-precision numerical propagator can be used, but then there is no need for ORBITECHO.

(3) Table 6 needs to be extended to the same models as table 1, being of particular interest the numerical propagator. Also please detail the force model used and the integrator.

Questions for authors

See weakness section.

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

Related papers

Discussion (0)

Sign in to comment.