acceptodds
Under review as a conference paper at ICLR 2027

Representation–Propagation Learning for Closed-loop Error-Controlled Long-Horizon Aircraft Trajectory Forecasting

Abstract

Long-horizon aircraft trajectory forecasting is challenging because uncertainty in future flight intent can be further amplified during physical rollout. We propose Representation–Propagation Learning (RPL), a forecasting framework that explicitly separates future-intent representation from trajectory propagation. RPL first infers a latent future intent and decodes it into a reference trajectory, which is then realized through differentiable aircraft dynamics with bounded closed-loop feedback. Our main theoretical result characterizes the sensitivity of this propagation process. By separating feedback-stabilized tracking variables from the unavoidable position integrators, we show that, under finite-gain closed-loop tracking dynamics, perturbations in future intent induce a terminal-state sensitivity that grows at most as O( √ T) with the forecast horizon, rather than exponentially. This provides a theoretical basis for controlling error amplification in long-horizon physical rollout without requiring contraction of the complete aircraft state. Experiments conducted on 21,816 held-out ADS-B forecasting windows demonstrate that RPL improves long-horizon forecasting accuracy and disturbance robustness in our evaluation. Although CFM performs better over short horizons, RPL’s advantages become apparent as the prediction horizon extends; at a 120-second prediction horizon, RPL reduces ADE and FDE by 9.79% and 5.17%, respectively, compared with CFM. Controlled perturbation experiments further show that the closed-loop feedback mechanism substantially reduces error accumulation and performance degradation during rollout, supporting RPL’s robustness in long-horizon trajectory prediction. You can get the code from this anonymous repository.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Aircraft trajectory forecasting, Representation–Propagation Learning, latent intent, differentiable dynamics, closed-loop stability, uncertainty propagation.

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_GqBbOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies long-horizon aircraft trajectory forecasting, focusing on how inaccuracies in an inferred future-motion hypothesis may be amplified during trajectory rollout. It proposes Representation–Propagation Learning (RPL), which separates future-intent representation from physical trajectory realization. A history encoder and reference decoder generate a reference future trajectory, which is subsequently realized through a differentiable reduced-order aircraft dynamics model with bounded control and closed-loop correction. The paper also analyzes propagation sensitivity by separating feedback-regulated tracking variables from position integrators. Experiments on 21,816 held-out ADS-B forecasting windows from TrajAir assess horizons of 30, 60, 90, and 120 seconds. The evaluation includes kinematic, recurrent, Transformer-based, and generative baselines, together with realization-module ablations and controlled-disturbance tests. RPL is less accurate than the CFM baseline at 30–60 seconds but performs better at 90–120 seconds. At a 120-second horizon, the reported ADE and FDE reductions relative to CFM are 9.79% and 5.17%, respectively.

Critical strengths

1. A clear separation between learned reference prediction and physical trajectory realization. The two-stage formulation is understandable (Sec. 2; Fig. 1). By evaluating the frozen reference, direct dynamics projection, and full realization module (Table 2), the paper offers a useful decomposition of where predictive improvements arise.

2. Long-horizon gains over a diverse set of baselines. The evaluation includes simple kinematic models, recurrent and Transformer-based predictors, and generative approaches (Table 1). Importantly, the paper reports the weaker short-horizon performance of RPL as well as its stronger results at 90–120 seconds, rather than presenting a uniformly favorable comparison.

3. Complementary ablation and disturbance experiments. The studies of reference smoothing, adaptive correction gating, realization training, and paired open-/closed-loop disturbances (Table 2; Figs. 4–5) help assess the roles of the main modeling components.

Critical weaknesses

1. Limited coverage and explanation of specialized aircraft-forecasting baselines. The main comparison includes ASCENT-adapted and CFM-Aircraft-adapted methods, but other relevant aircraft-forecasting approaches discussed in Appendix B.9, such as TrajAir-related interaction-aware predictors and the goal-oriented diffusion method GooDFlight, are not evaluated. The paper also lacks a direct comparison with a closely related physics-constrained or dynamically feasible forecasting method, despite physical realization being a central feature of RPL. The authors should better justify the selected baselines and explain the task adaptations and comparability of the aircraft-specific methods; additional comparisons would be welcome where the prediction settings are compatible.

2. Incomplete documentation of trajectory preprocessing and data partitioning. Section 4.1 indicates that detailed processing steps are provided in the appendix, but Appendix B.1 mainly reports the data source, sampling rate, state representation, historical/future window lengths, and split sizes. It does not specify the sliding-window stride, the handling of discontinuous or abnormal tracks, or whether data were partitioned by flight/day before window extraction. Since overlapping windows can share many observations, the partitioning procedure is important for interpreting the reported test results. This is an unresolved reporting question, not an assertion that data leakage occurred.

3. The main evaluation does not report several trajectory dimensions emphasized by the model. RPL produces a trajectory involving horizontal position, altitude, speed, heading, and vertical speed, yet Table 1 evaluates only horizontal ADE and FDE. The paper does not present a corresponding quantitative comparison of altitude error or physical/kinematic consistency across the main methods. These measures would help substantiate the practical value of its dynamics-aware realization stage beyond horizontal position accuracy.

4. The mapping between reference trajectories and abstract propagation inputs is not sufficiently specified. Sec. 3.1 and Appendix A.1 introduce reference-induced input sequences UU and VV and a tracking system Z˙=g(Z,U)\dot Z=g(Z,U), but do not explicitly define the mapping from a reference trajectory to these theoretical inputs or clarify how they correspond to the control commands and propagation law in Eqs. (3)–(4). Appendix B.4 describes reference-derived feed-forward commands, but does not establish their exact relationship to the abstract input UU. Clarifying this connection would make the proposed framework and its analysis easier to interpret.

Questions for authors
  1. Dataset construction: What exact window stride, discontinuity/quality filtering procedure, and train/validation/test partitioning unit were used? Were entire flights or collection periods separated before extracting windows?
  2. Aircraft-specific baselines: Why were ASCENT-adapted and CFM-Aircraft-adapted selected without additional comparison to the other aircraft predictors discussed in Appendix B.9? What task adaptations were required to make the reported comparisons meaningful?
  3. Trajectory-quality metrics: Can the authors provide altitude/vertical prediction errors and quantitative physical or kinematic-consistency metrics for RPL and the primary baselines?
  4. Reference-to-input mapping: What is the exact mathematical definition of the reference-induced input UU in Sec. 3.1? How is it constructed from the reference trajectory, and how does Z˙=g(Z,U)\dot Z=g(Z,U) relate to the feedback/control law and differentiable propagator in Eqs. (3)–(4)?
Reviewer_NogAOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The authors propose Representation-Propagation Learning (RPL), a framework for long-horizon aircraft trajectory forecasting that separates future-intent representation from physical trajectory propagation. Specifically, RPL first infers a latent trajectory-level intent from history and decodes it into a reference trajectory. Conditioned on the current state, this reference is then realized through a differentiable reduced-order aircraft model with bounded control and closed-loop feedback. The authors provide a finite-gain analysis of the propagation module, showing that under a tracking-dissipativity assumption and Lipschitz position dynamics, the propagation-induced terminal sensitivity grows at most as O(T)\mathcal{O}(\sqrt{T}) for fixed-energy input perturbations. Experiments on the TrajAir ADS-B dataset compare RPL against several neural and kinematic baselines across 30–120 s horizons, with ablations and controlled perturbation diagnostics.

Critical strengths
  1. This paper separates future-intent representation from physical realization, which is a reasonable and interpretable design.
  2. The theoretical claim is evidently supported: the O(T)\mathcal{O}(\sqrt{T}) result is explicitly for propagation-induced sensitivity, not total forecasting error. The derivation is understandable and the assumptions are stated.
  3. The paper compares against a wide range of baselines and reports results over multiple horizons and seeds.
  4. The component ablations and theory-aligned perturbation experiments provide useful insights on the significance of long horizon of the pipeline.
Critical weaknesses
  1. I think the novelty in this work is incremental. Combining a learned latent trajectory representation with a differentiable physics/control module and closed-loop feedback is not fundamentally new. The work is close to dynamically feasible trajectory forecasting, differentiable control, Neural ODEs, and control-based trajectory generation.
  2. While the theoretical analysis is neat, the main assumption is strong and is not guaranteed by the learned RPL system.
  3. RPL is worse than CFM at 30–60 s and only overtakes it at 90–120 s. At 120 s, the relative improvements are about 9.8% ADE and 5.2% FDE over CFM. These are not negligible, but they are not large.
  4. The state includes altitude, speed, heading, and vertical speed, but the reported metrics are horizontal ADE/FDE only. For an aircraft trajectory prediction paper, this seems to be a significant omission. The reference decoder predicts geometry and derives kinematics; the final evaluation should at least report altitude and vertical-rate errors.
  5. The physical model is limited. The differentiable reduced-order aircraft model is essentially a bounded kinematic integrator with feedback. It does not include aerodynamic forces, mass, thrust, or full six-degree-of-freedom dynamics.
  6. The paper’s central motivation is to separate representation error from propagation sensitivity. However, the experiments do not directly quantify how much of the final error comes from each source.
Questions for authors
  1. Can you comment on the primary assumption for the analysis and it's practicality?
  2. Can you include altitude and vertical-rate errors in the final results?
  3. What's the justification of using extremely simplified aircraft model?
  4. Can you disentangle representation error from propagation sensitivity?
Reviewer_6KpoOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes Representation–Propagation Learning (RPL) for aircraft trajectory forecasting. RPL first predicts a latent future intent from the observed history and decodes it into a reference trajectory, which is then realized through a differentiable reduced-order aircraft model with closed-loop feedback. A theoretical analysis shows that under a finite-gain assumption on the tracking dynamics, a propagation-induced terminal sensitivity that grows at most as O(√T) with the horizon. Experiments on the TrajAir ADS-B dataset show that RPL is worse than several neural baselines at 30–60 s but outperforms all baselines at 90–120 s. Perturbation experiments further show that closed-loop feedback reduces disturbance-induced error growth.

Critical strengths
  1. RPL presents an interesting framework for aircraft trajectory forecasting that combines a learned forecasting model with control methods, including future-intent representation and closed-loop physical propagation.
  2. The theoretical analysis of terminal sensitivity is clearly presented.
Critical weaknesses
  1. Presentation. Some symbols are used for different quantities. For example, K denotes the feedback policy in π_K but the number of prediction steps in K_H. In Section 2.2, it is hard to understand the meaning and structure of the notation.
  2. Theorical analysis seems interesting. But such close-loop propagation sensitivity seems well discussed in control field. It is unclear what is new or improved for the forecasting setting.
  3. Comparison with TrajAir-specific methods. The method proposed with the TrajAir dataset is not included as a baseline.
Questions for authors
  1. Do the authors report the number of parameters and the computational cost of RPL and the baselines? Is RPL smaller or more efficient than the neural baselines?
  2. Please clearly indicate which components of RPL are learned and which are fixed (e.g., the smoothing operator S, the feedforward command u^ff, the feedback policy π_K)
  3. Since the advantage of RPL grows with the horizon, does it continue at longer horizons?
  4. Please add background on the practical use of the forecasts: which applications (e.g., air traffic management or collision avoidance) require 90–120 s predictions, and is the weaker short-horizon performance acceptable for them?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.