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

1. Problem Statement

Accurate long-horizon vessel trajectory prediction remains a significant challenge in maritime intelligence due to the compounding uncertainty inherent in complex navigation behaviors and environmental factors. Existing approaches face two primary limitations:

1. Global Directional Inconsistency: Standard autoregressive models (e.g., Transformers, LSTMs) often fail to maintain global directional consistency over extended time horizons. As uncertainty accumulates, predictions tend to drift, collapse into implausible trajectories, or degrade into nearly horizontal paths.
2. Lack of Explicit Intent Modeling: Most current methods operate solely at the local kinematic level, implicitly inferring long-term navigational intent from short-term motion patterns. This becomes increasingly unreliable over time. Furthermore, token-based approaches that discretize continuous geographic space struggle to generalize to unseen regions, leading to unstable extrapolation.

The paper posits that increasing model capacity alone is insufficient; instead, the problem requires a hierarchical approach that explicitly separates high-level navigational decisions from low-level motion dynamics.

2. Methodology

The authors propose SKETCH, a semantic key-point conditioning framework that decomposes long-horizon prediction into global semantic decision-making and local motion modeling.

2.1. Core Concept: Next Key Point (NKP)

The framework introduces the Next Key Point (NKP) as a semantic intent variable. An NKP is defined as an extendable, predefined coordinate representing a semantically meaningful maritime node (e.g., port regions, straits, turning areas, or shipping-lane junctions) that a vessel is likely to approach.

• Role: NKPs act as semantic constraints that restrict the support of admissible future trajectories to semantically feasible subsets, rather than serving as precise geometric waypoints or final destinations.
• Factorization: The conditional distribution of the future trajectory $Y$ given history $X$ is factorized as:
  $$P(Y | X) = \sum_{Z} P(Y | X, Z) P(Z | X)$$
  where $Z$ represents the NKP. This decouples high-level intent inference ($P(Z|X)$) from conditional motion modeling ($P(Y|X,Z)$).

2.2. Model Architecture

The system utilizes a decoder-only Transformer (specifically a "MiniMind" architecture) adapted for continuous maritime data.

• Input Processing: Latitude and longitude are normalized to $[-1, 1]$, while Speed Over Ground (SOG) and Course Over Ground (COG) are converted to scaled velocity components.
• Physical Representation: To ensure numerical stability during autoregressive rollout, the model predicts SOG and COG. The next position is calculated using a locally Euclidean approximation of spherical motion (Proposition 3.3), avoiding repeated spherical projections and systematic error amplification.

2.3. Three-Stage Training Strategy

The learning procedure approximates the joint distribution $p(Y, Z | X)$ through three stages:

1. Conditional Trajectory Modeling (Blue):

   • The model is trained to predict $P(Y | X, Z)$ using "oracle" ground-truth NKPs as conditioning.
   • This stage focuses on learning fine-grained kinematic patterns under fixed semantic constraints.
   • To mitigate exposure bias, the authors employ a step-by-step training scheme combining Maximum Likelihood (GPT-style) and Behavior Cloning (imitation learning) losses.

2. NKP Modeling (Green):

   • The trajectory backbone is frozen. A lightweight Multi-Layer Perceptron (MLP) is trained to estimate $P(Z | X)$.
   • Contrastive Learning: The model uses a retrieval-augmented verification approach. Trajectories leading to the same NKP are pulled together in the embedding space, while those leading to different NKPs are pushed apart using a contrastive loss.
   • Open-Set Inference: Instead of a fixed closed set of ports, NKP inference is treated as an open-set problem. During inference, the model retrieves the most similar trajectories from a reference database to verify the predicted NKP, allowing generalization to unseen navigational targets.

3. Integrated Inference:

   • The system first estimates the NKP $\hat{Z}$ from the observed history $X$ using the trained NKP module.
   • The future trajectory is then generated by conditioning the trajectory predictor on $(X, \hat{Z})$.

3. Key Contributions

1. Identification of Intent Gap: The authors identify the lack of explicit global intent modeling as a fundamental limitation in existing long-horizon vessel trajectory prediction methods.
2. Semantic Conditioning: Introduction of the NKP as a semantic intent variable, separating global navigational decisions from local motion dynamics within a hierarchical framework.
3. Open-Set Generalization: An efficient training strategy for NKP-conditioned forecasting that allows the model to generalize to open-set navigational targets via retrieval-augmented verification, rather than relying on a fixed closed set of ports.
4. State-of-the-Art Performance: Demonstrated superior performance on real-world AIS data, particularly for long travel durations, directional accuracy, and fine-grained trajectory prediction.

4. Experimental Results

The method was evaluated on large-scale AIS datasets using three complementary metrics: Mean Squared Error of Position (MSEP), Mean Squared Error of Curvature (MSEC), and Mean Fréchet Distance (MFD).

• Overall Performance: SKETCH consistently outperformed state-of-the-art baselines (including MP-LSTM, TrAISformer, DiffuTraj, and AISFuser) across all metrics. It achieved the lowest MSEP (4.40E-5 vs. 8.47E-4 for MP-LSTM) and the smallest MFD (0.63 vs. 4.35 for MP-LSTM), indicating superior point-wise accuracy and global geometric alignment.
• Generalization: On a public AIS dataset from a different source (out-of-domain), SKETCH maintained the best performance, whereas TrAISformer suffered severe degradation due to its reliance on embedding layers that could not generalize to unseen coordinates.
• NKP Prediction: The open-set supervised fine-tuned model (sft-o-s2) achieved 95.46% accuracy in NKP prediction, significantly outperforming a Random Forest baseline (76.54%) and demonstrating robustness in selecting appropriate future maritime anchors.
• Ablation Studies:
  • NKP Quality: Using correct (oracle) NKPs yielded the best results. Incorrect NKPs caused substantial degradation, confirming that semantic guidance is critical but must be accurate.
  • Decoupling: Models that explicitly decoupled NKP representation from motion modeling outperformed those that injected NKP information directly into temporal models without decoupling.
  • Horizon Stability: As the prediction horizon increased (up to 144 points), SKETCH maintained stability, while other models showed consistent upward trends in error (MFD).

5. Significance and Claims

The paper claims that SKETCH provides a hierarchical framework that effectively reduces long-term uncertainty by explicitly modeling global navigational intent. By decoupling intent inference from conditional motion modeling, the approach improves trajectory coherence and prevents the drift common in autoregressive models.

The authors position this work not only as a solution for maritime navigation but as a general modeling paradigm for other long-horizon sequential prediction problems involving hierarchical decision structures. They suggest the approach serves as a foundation for downstream maritime AI tasks, including anomaly detection, route optimization, and multi-agent coordination, without making claims of immediate deployment in specific operational systems beyond the scope of the presented research.