Trajectory-Based Paradigm: Sequential Modeling

• Trajectory-Based Paradigm is a computational approach that models complete sequences to capture temporal continuity and spatial relationships.
• It employs techniques like motion pattern mining, Hankel matrix representations, and diffusion methods for effective analysis, prediction, and control.
• This paradigm offers enhanced robustness and interpretability over static, frame-wise methods, proving essential in autonomous driving, robotics, and urban mobility.

A trajectory-based paradigm refers to a family of computational, statistical, and algorithmic approaches that place the trajectory—a temporally ordered sequence of spatial or state observations—as the primary object of modeling, analysis, prediction, or control. Trajectory-based methods are especially prevalent in fields where sequential motion or event patterns are core to system understanding, such as autonomous driving, robotics, animal movement ecology, transport engineering, and networked systems. These approaches differ fundamentally from frame-wise, static, or aggregate methods by explicitly leveraging temporal continuity, sequential dependencies, or long-horizon relational information encoded in entire trajectories rather than isolated points or short fragments.

1. Foundational Principles of the Trajectory-Based Paradigm

Trajectory-based paradigms are characterized by two core principles: the explicit modeling of sequential evolution in a system’s state (spatial, velocity, or high-dimensional feature space), and the inference of behavior, prediction, or control from aggregate patterns across whole or partial trajectories.

Instead of treating data as unordered observations or as short temporal snippets, the paradigm considers the full continuity of motion or behavior encoded in the sequence:

• In trajectory clustering and motion pattern mining (Kalayeh et al., 2015), trajectories are decomposed into localized segments (flow vectors) and then recomposed into higher-level “motion patterns” via spatial, velocity, and reachability relationships.
• In system analysis and control, the trajectory-centric view treats all admissible system behaviors as lying within the vector space spanned by time-shifts of observed trajectories, facilitating direct data-driven system analysis without explicit model identification (Berberich et al., 2019).
• For prediction (e.g., animal or human movement), trajectory-based models forecast future paths by conditioning on the full historical sequence, often via deep sequence models, kernel methods, or diffusion models (Bae et al., 27 Mar 2024).

These approaches recognize that behaviors such as mode transitions, anomaly emergence, or coordinated flows are rarely visible in frame-wise or marginal statistics but are salient when considered through the entire sequence evolution.

2. Key Methodological Frameworks

Several distinct methodological archetypes crystallize within the trajectory-based paradigm, each with specialized instantiations:

Framework	Key Mechanism	Example Papers
Motion Pattern Mining	Flow vector decomposition, clustering	(Kalayeh et al., 2015)
Nonparametric Clustering	Dirichlet process, temporal gravity	(Kumaran et al., 2018)
Data-Driven System Control	Trajectory span/Hankel matrices	(Berberich et al., 2019)
Dynamic Graph Construction	Spatio-temporal cell partition, flows	(Kim et al., 2022)
Generative Sequence Models	GANs, diffusion, BNNs, transformers	(Jiang et al., 2023, Bae et al., 27 Mar 2024, Celestini et al., 31 Oct 2024)
Causal Representation	SCMs, backdoor adjustment	(Luo et al., 22 Apr 2024)
Curriculum RL for Control	Multi-agent, phased learning	(Garg et al., 31 Oct 2024)

For example, in the "motion pattern approach" (Kalayeh et al., 2015), the procedure is structured in four algorithmic phases:

1. Flow vector computation: Segmenting each trajectory into vectors that capture instantaneous movement—spatial location and velocity.
2. Motion component extraction: Clustering flow vectors into compact “motion components” via joint spatial/velocity distance (see the metric: $$D^2 = \Delta x^2 + \Delta y^2 + \beta(\Delta u^2 + \Delta v^2)$$).
3. Reachability set creation: Defining a motion component’s reachability based on spatial proximity (double-ellipse), flow alignment, and direction similarity criteria.
4. Motion pattern formation: Aggregating motion components into global patterns via agglomerative clustering, using weighted Jaccard distance over “signature” sets of reachably connected components.

Parallel developments—such as the TIGM/DEM family (Kumaran et al., 2018)—embed temporal dependencies into Bayesian mixture models, yielding unsupervised clustering that can identify both static motifs and their dynamically evolving variants.

Kernel, deep learning, and transformer-based frameworks extend trajectory-based inference to highly non-linear or high-dimensional settings, as in universal diffusion models for human dynamics (SingularTrajectory (Bae et al., 27 Mar 2024)) or sequential predictive controllers for robotic platforms (transformer-based MPC (Celestini et al., 31 Oct 2024)).

3. Theoretical Innovations and Mathematical Constructs

Trajectory-based paradigms are often formalized through new mathematical markers:

• Signature-based comparison: The use of motion component “signatures”—the set of all path-reachables—facilitates the quantification of global behavior similarity through advanced set metrics (e.g., weighted Jaccard distance in (Kalayeh et al., 2015)).
• Trajectory span representation: System trajectories $\{\bar{u}, \bar{y}\}$ can be entirely characterized via shifted sub-sequences in Hankel matrix form, allowing reconstruction of all valid behaviors by solving $[H_L(u)\ H_L(y)]\alpha = [\bar{u}; \bar{y}]$ (Berberich et al., 2019).
• Causal adjustment in trajectory spaces: To combat confounding, the structural causal model (SCM) is leveraged such that the effective representation is $P(H\,|\,do(X)) = \sum_{e} P(H\,|\,X, e)P(e)$, enabling disentanglement of environmental spurious signals (Luo et al., 22 Apr 2024).
• Diffusion process for trajectory generation: SingularTrajectory (Bae et al., 27 Mar 2024) unifies various tasks by first projecting trajectories into a low-rank “Singular space” and then iteratively refining trajectory candidates via diffusion-based denoising procedures, conditioned on both anchors and environmental traversability.
• BeLLMan residual minimization at trajectory scale: In reasoning with LLMs, the policy optimization objective transitions from stepwise to trajectory-level residuals, utilizing the model's own logits as soft Q-values for efficient off-policy value-based learning (Yuan et al., 21 May 2025).

These constructs allow fine-grained behavioral discrimination, invariant representation, and direct, interpretable use of data in modeling or control.

4. Applications and Impact

Trajectory-based paradigms support a broad class of applications by virtue of their attention to sequential patterning and temporal context:

• Traffic and transportation analysis: Identification of dominant movement patterns, lane utilization, merging/diverging behaviors, and anomaly/change detection (Kalayeh et al., 2015, Kim et al., 2022).
• Autonomous navigation and robotics: Precise trajectory selection for end-to-end planners, safety-critical performance through multi-candidate evaluation and rotation-based augmentation (Yao et al., 7 Jun 2025), and curriculum learning for robust tracking in cluttered aerial environments (Garg et al., 31 Oct 2024).
• System identification and control: Enable model-free simulation, robust controller synthesis, and direct inference for both linear and (lifted) nonlinear systems, bypassing traditional identification (Berberich et al., 2019, Celestini et al., 31 Oct 2024).
• Scene understanding and security: Anomaly detection in urban settings and surveillance applications by capturing deviations from learned typicality in spatio-temporal structure (Kumaran et al., 2018, Kim et al., 2022).
• Human movement and urban dynamics: Universal modeling of pedestrian/agent mobility, cross-domain transfer and few-shot adaptation, and robust causal reasoning about human-environment interaction (Bae et al., 27 Mar 2024, Luo et al., 22 Apr 2024).
• LLM reasoning: Value-based trajectory-level reinforcement learning for complex task-solving, with efficient credit assignment and off-policy optimization (Yuan et al., 21 May 2025).

Outputs in this paradigm have yielded demonstrable gains: e.g., TS-TrajGen achieves $>$97% reduction in divergence on macro-level human mobility metrics (Jiang et al., 2023), and TrajTrack reaches a new benchmark with a $4.48\%$ increase in tracking precision on nuScenes while running at $56$ FPS (Fan et al., 14 Sep 2025).

5. Comparison with Non-Trajectory Methods and Known Limitations

Trajectory-based methods contrast sharply with frame-wise or purely aggregate approaches:

• Frame-wise (two-frame) trackers deliver efficiency but lack robustness to long-term continuity, often failing in occlusion or sparsity (Fan et al., 14 Sep 2025).
• Sequence-based methods can address temporal depth but are computationally expensive when operating on raw high-dimensional data (e.g., point clouds).
• Regression-based trajectory prediction models only produce a single path, often miss safety nuances, and lack the discriminative power for rare, critical events (Yao et al., 7 Jun 2025).

The trajectory paradigm resolves these by compressing historical context (e.g., via bounding box trajectories), integrating global priors, and supporting adaptive, progressive refinement via candidate selection, self-distillation, or explicit fusion of fast local and learned global predictions.

However, challenges remain: accuracy of explicit prediction models (e.g., problems in Koopman-based velocity estimation (Tran, 28 Jun 2024)), the need for efficient fusion of explicitly and implicitly extracted motion cues, and ensuring adaptation to rapidly changing real-world conditions (e.g., traffic, environmental context).

6. Recent Advancements and Future Directions

Recent research highlights expansion into several new directions:

• Benchmarking and Evaluation Paradigms: Introduction of scenario-based benchmarking like CRITERIA (Chen et al., 2023) and scale/motion-aware ranking (Rasouli, 2023) for fair and robust trajectory evaluation.
• Causal trajectory representation: Explicit disentanglement of environmental confounding for improved generalizability and interpretability (Luo et al., 22 Apr 2024).
• Universal and transferable models: Unified embedding spaces (Singular spaces) and diffusion-based approaches for broad generalization across trajectory tasks (Bae et al., 27 Mar 2024).
• End-to-end planning: Coarse-to-fine safety-critical selection methods, rotation-based data augmentation, and robust refinements for autonomous driving (Yao et al., 7 Jun 2025).
• Direct code-driven trajectory adaptation: Natural language-LLM-driven modification of planner outputs by generating high-level plans and actionable code for robotic control, bypassing retraining or costly replanning (Maurya et al., 17 Apr 2025).

Looking ahead, promising research avenues include tighter integration of causal modeling, scalable multi-modal fusion (visual, LiDAR, language), and the development of trajectory-based pipelines with enhanced transparency, robustness to covariate shift, and cross-domain adaptability.

---

Trajectory-based paradigms mark a persistent shift toward temporally- and contextually-aware modeling in machine perception, control, and reasoning, offering unified frameworks that robustly capture, predict, and adapt sequential behaviors across diverse real-world domains.