SKETCH: Semantic Key-Point Conditioning for Long-Horizon Vessel Trajectory Prediction

AI summary

Overview

Research area: maritime robotics and intelligent transportation, specifically data-driven vessel trajectory forecasting from Automatic Identification System (AIS) data.

Technical level: Intermediate. The paper assumes familiarity with Transformer sequence models, autoregressive decoding, contrastive learning, and standard trajectory error metrics, but its central idea (conditioning on a "next key point") is conceptually simple.

Scope: The paper proposes a hierarchical, semantic-key-point-conditioned framework (SKETCH) for long-horizon vessel trajectory prediction and evaluates it against MP-LSTM, TrAISformer, DiffuTraj, and AISFuser on private and public AIS datasets.

What This Paper Is About

Long-horizon vessel trajectory prediction is hard because uncertainty compounds over time: models that only track local speed and heading tend to drift, flatten out, or produce physically implausible paths when extrapolated far into the future, and they have no explicit notion of where the ship is actually trying to go. The authors argue that this failure comes from the missing modeling of global navigational intent, since real vessels follow a hierarchy of high-level route planning plus low-level speed and heading control. Their goal is to restore that hierarchy by predicting an intermediate "Next Key Point" (NKP) such as a port, strait, turning area, or shipping-lane junction, and then generating the trajectory conditioned on that semantic target.

Key Contributions

1. The paper identifies the absence of explicit global intent modeling as a fundamental limitation of existing long-horizon vessel trajectory prediction methods.
2. It introduces the Next Key Point (NKP) as a semantic intent variable and conditions trajectory prediction on it, factorizing the problem into global navigational decision-making and local motion dynamics within a hierarchical framework.
3. It proposes a training strategy for NKP-conditioned forecasting that lets the model generalize to open-set navigational targets rather than depending on a fixed closed set of ports.
4. It reports experiments on large-scale AIS data showing state-of-the-art performance, particularly for long-horizon prediction.

Main Findings

• Best overall accuracy on the private test set (Table 2): the proposed model achieves MSEP 4.40e-5, MSEC 1.24e-6, and MFD 0.63, outperforming MP-LSTM (8.47e-4, 1.6e-5, 4.35), TrAISformer (0.35, 0.015, 12.95), DiffuTraj (0.0011, 1.6e-4, 3.33), and AISFuser (0.25, 0.0022, 20.51).
• Out-of-domain generalization holds (Table 3): on the public AIS_Dataset (Xie et al., 2025) with 560 samples and no training or fine-tuning by any model, the proposed method again leads with MSEP 2.81e-4, MSEC 1.37e-5, and MFD 0.97. TrAISformer degrades severely (MSEP 2.04, MSEC 0.038, MFD 49.50), which the authors attribute to its embedding layers failing on unseen latitude-longitude trajectories.
• NKP prediction is highly accurate (Table 4): pretrain-c reaches 98.98% closed-set accuracy, sft-o-s2 reaches 95.46% with open-set extension to unseen key points, sft-c reaches 92.32%, all above the Random Forest baseline (Zhang et al., 2020) at 76.54%. The open-set variant sft-o-s2 was chosen for integrated inference because it generalizes to new NKPs without retraining.
• NKP quality bounds downstream performance (Table 5): the oracle "correct NKP" setting gives MSEP 3.75e-5, MSEC 1.21e-6, MFD 0.61; the learned SFT-O is close at 4.48e-5, 1.69e-6, 0.63; SFT-c is 5.67e-5, 1.93e-6, 0.69; the adversarial "wrong NKP" setting degrades to 6.71e-4, 4.45e-4, 2.09; pretrain-c gives MSEP 5.18e-3, MSEC 3.20e-6, MFD 4.24; the NKP-free "pure 4ch" baseline gives 3.39e-4, 7.25e-4, 1.86. Learned NKP strategies sit between the oracle and adversarial bounds, and the wrong-NKP case is worse than having no NKP at all in MSEP and MFD.
• Consistent advantage across horizons (Table 7): from 12 to 144 prediction points, MFD rises monotonically for every method, but the proposed model stays lowest throughout, going from 0.012 at 12 points to 0.631 at 144 points, versus MP-LSTM (0.310 to 4.347), TrAISformer (5.090 to 12.949), DiffuTraj (1.980 to 3.327), and AISFuser (2.009 to 20.509). DiffuTraj and AISFuser are marked as closed-source and were implemented according to their original papers.
• Numerical consistency of the coordinate update: a one-step displacement check on real AIS trajectories of segment length 288, comparing the next-step position derived from SOG/COG against ground truth, yields a mean squared error on the order of 10^-9.
• Qualitative behavior: the authors report that MP-LSTM relies on only two anchor points (support and destination) and can deviate globally if either is wrong, while TrAISformer shows pauses and stagnant latitude; the proposed model is described as capturing both straight and turning trajectories with correct timing and coordinates.
• Not reported: the numeric values of Table 6 (the 4ch / 6ch / 6ch with wrong-NKP ablation) are not included in the available paper text, and inference duration values are described as a metric but no timing numbers appear in the excerpt.

Methodology in Plain English

The authors start from a probability statement: instead of directly modeling the future trajectory given the past, they write it as a sum over possible next key points, where each key point defines a plausible future and the model also estimates how likely each key point is. Because conditioning on a key point shrinks the space of admissible futures, the trajectory predictor only has to solve a smaller, better-posed problem.

Training happens in three stages. First, the trajectory predictor is trained with the ground-truth key point supplied as an oracle, so it learns motion patterns under a fixed semantic constraint; this uses a MiniMind decoder-only Transformer with linear input and output projections, normalizing latitude and longitude to [-1, 1] and turning course over ground (COG) and speed over ground (SOG) into scaled velocity components with a factor of 1/25. The prediction is autoregressive, factorized step by step, and trained with a step-by-step scheme that alternates a velocity-style loss with a coordinate-style behavior-cloning loss to reduce error accumulation. Second, the trajectory backbone is frozen and only a lightweight MLP is fine-tuned to predict the key point, using a contrastive loss (with a margin) that pulls together trajectories leading to the same key point and pushes apart those leading to different ones, following a retrieval-augmented verification idea in which an unknown trajectory is matched against a database of reference trajectories by cosine similarity. Third, at inference, the key point is predicted first and the trajectory predictor is conditioned on that estimated key point.

To keep the geometry stable over long rollouts, the model assumes locally linear motion (constant SOG and COG over a short interval) and uses closed-form latitude and longitude update equations involving the Earth's radius, avoiding repeated spherical projections.

Why This Matters

Impact on research: the paper reframes long-horizon trajectory prediction as a hierarchical intent-plus-motion problem rather than a pure sequence-modeling scaling problem, and it argues that tokenizing continuous geographic space is a source of failure for models like TrAISformer and TrackGPT when deployed outside their training regions. It also provides an information-theoretic and decision-theoretic justification for why conditioning on an auxiliary variable reduces uncertainty even though posterior probabilities are not improved pointwise.

Real-world applications named or implied by the paper:

• Collision avoidance at sea.
• Port operation optimization.
• Search and rescue.
• Fuel-efficient voyage planning.
• Strategic maritime transportation analysis, including voyage-status estimates for ship operators and cargo owners.
• Long-term route and schedule assessment under varying environmental conditions.

Industry relevance: maritime transport carries over 80% of goods worldwide by volume, and the widespread adoption of AIS has created large-scale, high-frequency vessel motion records. The authors note the work is based on civil container vessel data and that the data contain historical vessel trajectories with no personal or sensitive information. Code is released at https://github.com/LinyongGAN/SKETCH.

Future Directions

• Extending the open-set NKP database incrementally as new ports, straits, or lane junctions emerge, so long-horizon prediction keeps pace with an evolving maritime landscape without retraining.
• Applying the intent-conditioned factorization to other long-horizon sequential prediction problems that have hierarchical decision structures, which the authors explicitly propose as a general paradigm.
• Building downstream maritime AI capabilities on the learned representation, including anomaly detection, route optimization, and multi-agent coordination.
• Reducing the residual gap between learned NKP prediction and the oracle-NKP upper bound, since the ablation shows NKP errors cause measurable but bounded degradation and wrong NKPs can be worse than no NKP at all.
• Testing whether the tendency of token-based models to collapse toward nearly horizontal trajectories can be addressed by semantic conditioning rather than by increasing model capacity.

Target Audience

Researchers and practitioners working on maritime intelligence, vessel traffic services, and autonomous navigation; machine learning engineers interested in hierarchical or intent-conditioned sequence forecasting; and readers of robotics and transportation venues who want to see how semantic intermediate variables can stabilize long-horizon prediction from AIS data.

Authors’ abstract

Accurate long-horizon vessel trajectory prediction remains challenging due to compounded uncertainty from complex navigation behaviors and environmental factors. Existing methods often struggle to maintain global directional consistency, leading to drifting or implausible trajectories when extrapolated over long time horizons. To address this issue, we propose a semantic-key-point-conditioned trajectory modeling framework, in which future trajectories are predicted by conditioning on a high-level Next Key Point (NKP) that captures navigational intent. This formulation decomposes long-horizon prediction into global semantic decision-making and local motion modeling, effectively restricting the support of future trajectories to semantically feasible subsets. To efficiently estimate the NKP prior from historical observations, we adopt a pretrain-finetune strategy. Extensive experiments on real-world AIS data demonstrate that the proposed method consistently outperforms state-of-the-art approaches, particularly for long travel durations, directional accuracy, and fine-grained trajectory prediction.