Wind as Remapping and Phase Drift: A Map-Anchored Model of Terminal-Airspace Trajectories

Small aircraft landing at an uncontrolled airport do not fly arbitrary curves. They fly a rectangular "traffic pattern" fixed to the runway. Wind changes their paths in two different ways. First, it decides which runway end is in use, which flips the whole pattern. Second, within a pattern, it speeds aircraft up or slows them down along each leg, because pilots hold a ground-referenced track rather than drifting sideways. Current trajectory predictors treat wind as one more feature to feed in, or leave it out. We propose to build that two-way split directly into the model. The idea is borrowed from how the rodent hippocampus handles context. A discrete context signal selects which stored spatial map is active ("remapping"). A continuous velocity signal moves the animal's estimated position along that map ("path integration").

In two sentences: the model keeps a small library of learned route shapes tied to the runway, and represents each aircraft as a position (a "phase") along one of those routes. Wind and other external variables are allowed to do only two things: pick the route, and change how fast the phase advances. We expect this restriction to generalize to wind conditions rarely seen in training better than models that absorb wind into an unconstrained encoder. We test this on real terminal-airspace data with wind regimes held out of training.

Core Claim

External ("exogenous") variables affect navigating agents through two separate channels. One is a discrete choice of map. The other is a continuous change in the rate of progress along that map. A forecaster that hard-codes this split should beat an equally sized forecaster with unconstrained wind conditioning. The advantage should show up most on held-out wind regimes, such as runway reversals and strong crosswind or tailwind days, and least on in-distribution test days.

This claim differs from the obvious combination of the two research directions. The research file's leading gap is a path integrator whose velocity input is corrected by a wind field, roughly ground velocity = air velocity + wind. That version has two problems. The available data cannot support it: TrajAir and TartanAviation provide only a single METAR wind vector per weather report, not a spatial wind field. It also reduces to textbook vector addition, which any kinematic decoder can absorb. Our claim is about structure instead: which part of the trajectory wind is allowed to touch.

Prior Work and Its Role Here

The neuroscience side supplies the structure. A theory of joint attractor dynamics in the hippocampus and the entorhinal cortex couples velocity-driven grid-like modules to a hippocampal network that stores several environment maps. In that model, feedback from the stored map limits accumulated drift, and switching context switches the active map. We reuse this two-level organization. Unlike that work, our velocity input is deliberately not equal to true displacement, and our map switch is driven by an external variable rather than by artificial stimulation. What do grid cells contribute to place cell firing? and Learning place cells and remapping by decoding the cognitive map motivate anchoring the map to fixed geometry, which here is the runway, instead of relying only on integrated motion. Back to the Continuous Attractor constrains the design. It shows that learned continuous attractors in practice become slowly drifting manifolds whose short-term error is bounded by the size of the perturbation. That is why we let a learned phase manifold carry the trajectory, but expect a residual drift that the map anchor must absorb. Prediction and generalisation over directed actions by grid cells shows that actions can act as shifts on a fixed basis. Its reliance on translation invariance is exactly what fails near a runway, so we use a learned route manifold rather than a Fourier lattice.

The trajectory side supplies the testbed and the competitors. Predicting Like A Pilot introduces TrajAir and TrajAirNet, which encodes METAR wind with a CNN alongside trajectory and social encoders inside a conditional variational autoencoder (a generative model that samples futures from a learned latent distribution). That is the clearest instance of the "wind as a feature" approach. ASCENT normalizes coordinates to the aircraft's body frame, embeds runway landmarks, and decodes physical controls through a kinematic layer. It omits wind entirely. Our model keeps ASCENT's physical decoding instinct and adds a structured place for wind. From the time-series literature, TimeXer embeds each external variable as a single token joined by cross-attention. It reports robustness to missing exogenous data but trouble with sudden spikes caused by a single-token bottleneck. We adopt its variate-token encoder as the fair "unconstrained" wind conditioner. Aura and ST-ResNet (Zheng et al. (2017)) both report that separating exogenous streams beats naive concatenation. That motivates separation, but neither restricts what the exogenous stream may change, which is our point of departure.

Research Program

The model has three parts.

The first is a small library of $K$ route manifolds, learned from data in runway-anchored coordinates. Each manifold $r_k(\phi)$ is a smooth closed or open curve, for example a periodic spline. Here $k$ is the index of the route and $\phi \in [0,1)$ is the position along it. These play the role of stored hippocampal maps: traffic-pattern variants for each runway direction, straight-in approaches, and departures.

The second is a gate. It reads the recent trajectory and the METAR wind token and outputs a distribution over $k$. This is the remapping channel.

The third is a phase integrator. It advances the position along the active route according to

$$\frac{d\phi}{dt} = \frac{s(h) + \alpha_k(\phi)\, \mathbf{w} \cdot \hat{\mathbf{t}}_k(\phi)}{\lVert \partial r_k / \partial \phi \rVert},$$

where:

  • $s(h)$ is a learned airspeed term computed from the aircraft's history $h$;
  • $\mathbf{w}$ is the METAR wind vector;
  • $\hat{\mathbf{t}}_k(\phi)$ is the unit tangent of route $k$ at phase $\phi$;
  • $\alpha_k(\phi)$ is a learned gain, initialized at 1 and allowed to depart from it.

Projecting wind onto the route tangent encodes the pilot-correction prior: the along-track wind component changes groundspeed, and the cross-track component is cancelled by crabbing. A small residual network predicts the lateral offset from the route and is penalized to stay small. The predicted position is $r_k(\phi) + \text{offset}$. Sampling over $k$ and over a noise term in $s$ gives multimodal futures, which matches the min-of-K evaluation used in this literature.

The analysis has two strands. The first asks whether the learned gate recovers the runway in use from wind direction without being told. The second asks whether the learned gain stays near the physical value of 1 or departs from it systematically, for example on base and final legs, where pilots also change airspeed. These analyses tell us whether the structure is doing physical work or merely acting as a regularizer.

Baselines

The naive reference point is constant-velocity extrapolation of the last observed ground velocity. TrajAirNet's paper reports beating constant-velocity extrapolation by a wide margin. Any model that fails to beat it clearly is broken.

The first published competitor is TrajAirNet Predicting Like A Pilot. It is the only cited model that already conditions on METAR wind. Its paper reports a 120-second Average Displacement Error (the mean distance between predicted and true positions over the horizon) of "~0.71 to 0.86 km" on TrajAir. It reports this as 30–50% lower than Transformer and STG-CNN baselines. We will retrain it on our splits from the released code path and report its numbers on our protocol alongside the published figure.

The strongest existing method is ASCENT ASCENT. It reports a Final Displacement Error (the distance at the last predicted step) more than 50% lower than TrajAirNet, and 17 ms batch latency against TrajAirNet's 673 ms. The file does not give an absolute ASCENT FDE: no reported result found on Lacuna for ASCENT absolute ADE/FDE on TrajAir or TartanAviation. We therefore reimplement it and report its numbers ourselves. Because ASCENT ignores wind, comparing against it alone would be unfair in our favor. We therefore also build ASCENT+wind, which adds a TimeXer-style variate token for wind and visibility, joined by cross-attention to the motion encoder. This variant is the main bar.

All models get the same inputs, the same history length, the same 120-second horizon, and the same number of samples K. Parameter counts will be matched within 20%, and every model gets the same hyperparameter search budget, for example 30 trials on the validation split. Beating ASCENT+wind means two things on held-out wind regimes. First, a statistically reliable reduction in min-of-K ADE and FDE, with confidence intervals bootstrapped over days rather than over samples. Second, no in-distribution loss larger than a few percent.

Three ablations isolate the key ingredient:

  • Remap-only: wind drives the gate, and the phase speed ignores wind.
  • Drift-only: a single gate that ignores wind, with wind entering the phase speed.
  • Free-wind: the same route-manifold model, but wind fed into an unconstrained network that can move the offset as well as the phase.

If free-wind matches the full model, the restriction is not what helps. In that case the gain would come from the route library alone.

Why not just…?

Why not just add the wind to the velocity in ASCENT's kinematic layer? This vector-addition fix is physically cheap and is the real threat to a naive version of this proposal. For ground-referenced flight, though, it pushes the predicted track sideways, while pilots crab to stay on the pattern. The file offers no evidence either way on TrajAir, so we include it as an extra baseline (ASCENT+wind-triangle). If it matches our model on held-out crosswind days, the drift channel is unnecessary and the contribution shrinks to the remapping gate, which we would then report as such.

Why not just give a strong encoder the wind and let it learn? TrajAirNet does roughly this with a CNN, and ASCENT+wind does it with cross-attention. Prior exogenous work (Aura, TimeXer) shows that separating the streams beats concatenation within the training distribution. That evidence says nothing about rare regimes such as runway reversals, where an unconstrained encoder has few examples to learn from. The held-out-wind split is built to settle exactly this. If the gap between our model and ASCENT+wind is no larger out of distribution than in distribution, the claim fails.

Why not just use a mixture of experts gated by wind? This is essentially our remap-only ablation with neural experts in place of route manifolds. If it matches the full model, the continuous phase-drift channel adds nothing measurable, and the honest conclusion is that wind matters only as a context selector.

Proposed Experiments

The primary data is TrajAir at KBTP, which has 111 days of ADS-B data with synchronized METAR Predicting Like A Pilot. We make three splits. The first is a random-day split, for in-distribution performance. The second holds out runway-reversal days: all days on which the less common runway direction is active are reserved for testing, determined from METAR wind direction, and the model is trained only on the other direction plus a small number of reversal days. The third holds out strong winds, defined as the top decile of wind magnitude.

The secondary data is TartanAviation TartanAviation, which has 661 days covering the non-towered KBTP and the towered KAGC. We use it for scale and for a transfer test. Route manifolds are relearned at KAGC from a few days of data while the gate and phase-speed networks are kept frozen. If wind's effect really is structural, those components should transfer. We also report results separately for the low-visibility and precipitation subsets the dataset identifies.

The metrics are min-of-K ADE and FDE at 120 seconds, the accuracy of the gate's route choice against runway-in-use labels derived from the trajectories, and the coverage of 95% prediction regions.

The decisive checks are two.

The first is the generalization gap: the full model's improvement over ASCENT+wind and ASCENT+wind-triangle must be larger on the reversal and strong-wind splits than on the random split. If it is not, the structural claim is falsified.

The second is physical consistency. In the full model, the learned phase speed on each leg should track the along-leg headwind component, with a fitted gain near 1 on crosswind and downwind legs. A gain far from 1 everywhere would mean the drift channel is acting as a generic regularizer, not as the mechanism we propose.