PaperlayerView exact revision on arXiv

Source-linked AI summary

Trajectory Prediction for Autonomous Driving: Progress, Limitations, and Future Directions

Nadya Abdel Madjid, Abdulrahman Ahmad, Murad Mebrahtu, Yousef Babaa, Abdelmoamen Nasser, Sumbal Malik, Bilal Hassan, Naoufel Werghi, Jorge Dias, Majid Khonji

arXiv:2503.03262v3cs.ROcs.AIcs.CVcs.LG

TL;DR

Safe autonomous driving requires forecasting the trajectories of diverse, interacting traffic agents, but existing methods and evaluation settings remain fragmented. The paper reviews these methods through a comprehensive taxonomy and pipeline overview, then synthesizes active research areas, empirical findings, and unresolved limitations. It concludes that trajectory prediction remains an open research problem requiring broader evaluation, uncertainty handling, and deployment-aware development.

  • Problem

    Trajectory prediction must support safe navigation despite diverse agents, interactions, multimodality, and uncertainty, while the field lacks consistent ways to compare its many methods.

  • Method

    The paper conducts an in-depth review that classifies learning-based and non-learning methods and surveys prediction inputs, outputs, datasets, metrics, paradigms, and active research directions.

  • Results

    Map information improves displacement-error results in reported cases, including VectorNet’s 28% reduction at a 3-second horizon and LaneGCN’s reduction from 1.90 to 1.35 meters at 6 seconds.

  • Takeaways & Limitations

    The review identifies uncertainty estimation, driving knowledge, planning-conditioned prediction, VLM guidance, vehicle collaboration, and standardized cross-dataset evaluation as important directions.

  • Takeaways & Limitations

    Current research is constrained by isolated dataset evaluations and the lack of unified, interpretable parametrization of traffic scenarios for fair cross-dataset comparison.

Abstract

from arXiv · show

As the potential for autonomous vehicles to be integrated on a large scale into modern traffic systems continues to grow, ensuring safe navigation in dynamic environments is crucial for smooth integration. To guarantee safety and prevent collisions, autonomous vehicles must be capable of accurately predicting the trajectories of surrounding traffic agents. Over the past decade, significant efforts from both academia and industry have been dedicated to designing solutions for precise trajectory forecasting. These efforts have produced a diverse range of approaches, raising questions about the differences between these methods and whether trajectory prediction challenges have been fully addressed. This paper reviews a substantial portion of recent trajectory prediction methods proposing a taxonomy to classify existing solutions. A general overview of the prediction pipeline is also provided, covering input and output modalities, modeling features, and prediction paradigms existing in the literature. In addition, the paper discusses active research areas within trajectory prediction, addresses the posed research questions, and highlights the remaining research gaps and challenges.

1. Introduction

Trajectory prediction is essential for safe autonomous driving, yet diverse traffic participants, interactions, rule violations, multimodality, and operational constraints make accurate forecasting difficult. This review organizes existing solutions, surveys the prediction pipeline and modeling approaches, and identifies active research areas and remaining gaps.

  • Motivation: Accurate prediction of surrounding traffic trajectories supports autonomous-driving planning by helping eliminate collision-prone paths and select safer actions.The forecasting problem is complicated by varied scenarios, road layouts, agent interactions, maneuvers, human randomness, and computational requirements.
  • Prediction challenges: Vehicle and pedestrian trajectories depend on past motion, road topology, physical constraints, nearby-agent behavior, social norms, intent, and behavioral randomness.Models must represent each target agent’s surroundings and the interdependencies among traffic participants’ decisions.
  • Prediction challenges: Traffic-rule compliance narrows plausible futures, but real-world predictors must also account for violations such as speeding, illegal lane changes, and running red lights.The same past trajectory can also lead to multiple plausible futures, making trajectory prediction inherently multimodal.
  • Industry context: Industry systems balance predictive accuracy with real-time execution, cost-effectiveness, uncertainty handling, generalization, sensor choices, and deployment constraints.Tesla is described as vision-based and end-to-end, whereas Waymo uses a modular stack combining LiDAR, cameras, and radar.
  • Review scope: The review surveys existing trajectory-prediction techniques and provides a structured framework for comparing modeling approaches and identifying research gaps.Its research questions address existing solutions, whether the problem remains open, and emerging research trends.
  • Contributions: The paper covers the prediction pipeline, including input modalities, modeling features and approaches, output modalities, evaluation metrics, and prediction paradigms.It also extends existing taxonomies with comparative tables and reviews uncertainty, driving knowledge, planning-conditioned prediction, VLM-guided models, and vehicle collaboration.

2. Overview of Prediction Pipeline

This section presents a general trajectory-prediction pipeline, from problem formulation and inputs through modeling, outputs, evaluation, datasets, and research gaps.

  • The pipeline covers problem formulation, input modalities, modeling features and approaches, output modalities, evaluation metrics, benchmark datasets, and existing gaps.

2.1. Problem Formulation

Trajectory prediction estimates future state sequences for traffic agents from historical observations, with optional contextual inputs and coordinate representations.

  • Each traffic scenario contains an ego vehicle and agents represented by state vectors over an observation period.The state vector can minimally encode position and may additionally include orientation and speed.
  • The prediction task estimates each agent’s future trajectory over a specified horizon from previous states.
  • The predictor can use past trajectories alone or incorporate road geometry and interaction features as additional inputs.
  • Trajectories may be represented in local or global coordinates, although global-frame prediction requires precise ego-vehicle localization.
  • Frenet coordinates encode motion by arc length along the lane centerline and lateral offset, helping separate motion patterns from global scene geometry.

2.2. Input Modalities

Trajectory prediction systems use past trajectories, sensor data, and HD maps as complementary input modalities. These choices involve trade-offs among contextual richness, robustness, localization dependence, computational cost, and sensor limitations.

  • Input modalities comprise past agent trajectories, LiDAR, radar, cameras, and HD maps describing static environmental context.
  • Combining past trajectories with surrounding-agent histories captures interactions, but single-source models trained on error-free observations can degrade with noisy or incomplete real-world inputs.
  • LiDAR provides lighting-invariant perception and 360° situational awareness, but heavy rain, snow, reflective surfaces, cost, and fine-grained pedestrian recognition constrain deployment.
  • BEV rasterization transforms sensor data into a top-down representation aligning positions, velocities, semantics, and temporal information in a unified plane.
  • HD maps can improve forecasting in structured environments: VectorNet reduces 3-second displacement error by 28%, while LaneGCN reduces 6-second error from 1.90 to 1.35 meters.
  • HD-map pipelines require precise localization and ongoing map construction, while combining maps with sensors adds preprocessing that affects inference time.

2.3. Modelling Features

Modeling features describe static scene structure, agent state, semantic attributes, and interactions among traffic participants. Their usefulness depends on integration quality, data diversity, annotation availability, and inference cost.

  • Static features such as road geometry, sidewalks, and walkable areas impose spatial constraints that help trajectories align with the scene layout.
  • Dynamic features include past trajectories, velocity, heading, orientation, and object category, enabling models to learn motion and category-specific dynamics.
  • Semantic attributes can improve contextual representation, but limited annotations, propagated extraction errors, and added inference time complicate their use.
  • Interaction Features: Social pooling aggregates neighboring-agent features, commonly through max or average pooling, to represent interactions in a compact form.
  • Interaction Features: Interaction-aware performance depends on integration design and scenario diversity; AgentFormer ablations show degraded performance when social or temporal cues are modeled alone.

2.4. Modelling Approach

Trajectory prediction approaches range from interpretable model-based methods to sampling-based, diffusion, and learning-based methods. The appropriate choice depends on application requirements, platform characteristics, environmental complexity, and desired handling of long-horizon dynamics.

  • Model-based methods such as Constant-Velocity and Interactive Multiple Model explicitly encode motion constraints, producing physically plausible trajectories with simple implementation and interpretation.
  • Data-driven methods are commonly used to capture complex trajectory patterns over longer horizons by modeling spatial and temporal context.
  • Diffusion models add Gaussian noise and iteratively denoise from random initialization while conditioning predictions on past trajectories and scene context.
  • Classical machine-learning approaches estimate probabilistic parameters from data, whereas deep-learning methods learn hierarchical and nonlinear representations for complex dynamic settings.
  • Interaction modeling and intention-awareness incorporate relationships among agents and planning-based reasoning into trajectory prediction.

2.5. Output Modalities

Trajectory prediction outputs may represent agent-centered futures as unimodal or multimodal trajectories, maneuver intentions, or occupancy probabilities. These formats differ in how they express alternative behaviors and support downstream planning.

  • Object-centric outputs are categorized as unimodal or multimodal, with unimodal models producing deterministic points or parameters of a unimodal distribution.
  • Multimodal outputs represent alternative plausible futures, such as proceeding straight or turning left from a lane permitting both maneuvers.
  • Maneuver outputs classify intentions such as lane changes, lane keeping, going straight, and turning left or right.
  • Joint trajectory-and-maneuver prediction uses either multi-task outputs with shared feature extraction or related joint modeling strategies.
  • Occupancy-centric models estimate the likelihood that each grid cell will be occupied at future time steps, with grid resolution balancing precision and computational demand.

2.6. Performance Evaluation

Trajectory prediction evaluation uses displacement, probabilistic, matching, open-loop, closed-loop, and deployment-oriented measures. Fair comparisons also require accounting for object identity, sampling rates, downstream effects, and inference time.

  • Open-loop evaluation directly compares predicted trajectories with ground-truth future paths, but does not quantify downstream effects on other modules.
  • ADE and FDE measure average L2 displacement over the full prediction horizon and at its final time step, respectively.
  • MinADE_k and MinFDE_k extend displacement metrics to multimodal predictions by selecting the closest of k candidate trajectories.
  • NLL evaluates probabilistic predictors by measuring the likelihood assigned to the ground-truth trajectory under the predicted distribution.
  • When ground-truth identifiers are unavailable, predicted and ground-truth objects must be matched before accuracy metrics can be calculated.
  • Closed-loop evaluation captures prediction-error propagation into planning and ego-vehicle safety, while inference time reflects suitability for real-time deployment.
  • Cross-dataset comparisons require resampling when training and deployment sampling rates differ, because sampling determines the learned and deployed motion dynamics.

2.7. Datasets and Leaderboard

Trajectory prediction benchmarks span diverse environments, sensors, traffic conditions, and prediction horizons. The surveyed datasets also differ in standardization, map resources, annotation formats, and evaluation protocols.

  • Benchmark scenarios include intersections, highways, mixed urban streets, crowded public spaces, and residential areas to test varied complexity, location, and traffic density.
  • KITTI lacks an official trajectory-prediction split, motivating a defined split with 8,613 top-view trajectories for training and 2,907 for testing.
  • KITTI-360 covers 70 kilometers with OpenStreetMap geolocation, while the Daimler dataset targets pedestrian prediction over horizons shorter than 2 seconds.
  • ApolloScape provides camera and LiDAR trajectories from Beijing under varied lighting and complex traffic, with prediction evaluated over the next 3 seconds.
  • The Waymo Motion Prediction Challenge contains around 100,000 20-second scenes with detailed 3D bounding boxes and high-definition maps.
  • Table 2 organizes benchmark datasets for autonomous-driving trajectory prediction, including resources with diverse scene durations, tracked agents, and map representations.

2.8. Prediction Paradigms

The survey distinguishes modular, jointly optimized, and query-based prediction paradigms. These approaches trade debugging and modularity against error propagation, integration complexity, computational overhead, and input-data requirements.

  • Detect-Track-Predict: Detect-Track-Predict sequentially detects objects, maintains tracklets, and forecasts future motion from active tracks.
  • Detect-Track-Predict: Its modular design supports debugging and benefits from extensive detector and tracker research, but sequential errors propagate into degraded predictions.
  • Detect-Track-Predict: LiDAR-camera fusion can occur early, late, or in a unified bird’s-eye-view representation, with BEVFusion reporting 190 ms detection inference time.
  • Joint Learning: Joint learning optimizes detection and trajectory prediction under a unified objective, using pooled object features or shared scene representations to connect the tasks.
  • Joint Learning: Joint-learning designs vary in gradient sharing: PnPNet allows prediction loss to propagate through the detector and encoder, whereas FaF uses a shared backbone without full gradient sharing.
  • Query-Based Prediction: Query-based models use spatial and temporal object queries that attend to scene and map features, jointly producing detections and forecasts while reducing subsystem error propagation.
  • Query-Based Prediction: Query-based methods support holistic scene understanding but may incur computational overhead in dense scenarios and depend on accurate scene encoding, HD maps, and LiDAR.

3. Survey of Modelling Methods

The survey classifies trajectory prediction by how motion is modeled, distinguishing physics-based methods from learning-based approaches and their major subfamilies. It compares their modeling strengths, computational properties, uncertainty handling, and limitations across dynamic driving settings.

  • Taxonomy: The taxonomy separates physics-based methods, which define motion equations, from learning-based methods, which learn motion patterns from observations.Learning-based methods are further divided into classical machine learning, deep learning, and reinforcement learning.
  • Physics-based methods: Physics-based approaches include direct dynamic or kinematic models, Kalman filtering, and Monte Carlo sampling for state prediction and uncertainty modeling.Direct models assume known motion equations, Kalman filters represent Gaussian state uncertainty, and Monte Carlo methods approximate distributions through sampling.
  • Learning-based methods: Classical machine learning methods use probabilistic function approximators, while deep learning methods model complex hierarchical patterns with sequential, convolutional, graph, transformer, autoencoder, and adversarial architectures.The survey reviews Gaussian processes, hidden Markov models, dynamic Bayesian networks, Gaussian mixture models, and several deep learning model families.
  • Learning-based methods: Reinforcement learning methods learn from expert demonstrations and can produce multimodal trajectories under partial observability, but training is costly and learned rewards are difficult to formalize and interpret.Inverse reinforcement learning must represent safety constraints, social dynamics, and ambiguous traffic-agent intent.
  • Physics-based methods: Physics-based models are computationally efficient for short-term prediction and abrupt maneuvers but have limited motion-history flexibility and may require scenario-specific tuning.Constant-motion models omit uncertainty, Kalman variants impose Gaussian noise assumptions, and Monte Carlo methods are more expressive but computationally demanding.
  • Learning-based methods: Classical machine learning offers interpretability, uncertainty estimates, and performance with limited data, but limited parameterization restricts complex behavior modeling and adaptation to new environments.Unlike deep models, these methods cannot learn hierarchical representations effectively.
  • Learning-based methods: Deep learning strengths and limitations depend on architecture: GNNs model dense interactions, GANs generate diverse trajectories, and dynamic graphs require consistent object identities across frames.LSTMs capture temporal dependencies but struggle with spatial reasoning and long-range dependencies; GANs require substantial data and may fail under distribution shifts.

4. Incorporating Driving Knowledge

Driving knowledge is incorporated into trajectory prediction to represent environmental structure, agent behavior, and traffic interactions. The survey organizes this knowledge into map-related, agent-related, and traffic-related categories and reviews corresponding encoding strategies.

  • Motivation: Formalized driving knowledge can improve prediction performance and explainability by combining learned motion patterns with road, vehicle, and traffic constraints.The paper emphasizes that modular knowledge should remain adaptable to the deployment region.
  • Knowledge categories: The survey groups driving knowledge into map-related information, agent-related information, and traffic-related knowledge.These categories cover static road structure, agent kinematics and behavior, and traffic rules and interaction patterns.
  • Map-related knowledge: Map encodings represent road geometry, traffic controls, crossings, and signs, with vectorized graph representations preserving structural and semantic details better than rasterized maps.VectorNet uses polyline-based local and global graphs to encode lane and actor-map interactions.
  • Agent-related knowledge: Agent-related knowledge incorporates vehicle kinematics, dynamics, driving styles, and pedestrian mobility constraints to restrict predictions according to feasible motion.Examples constrain turning radius, acceleration, heading, and other vehicle-dynamic outputs.
  • Traffic-related knowledge: Traffic-related knowledge encodes rules and interaction patterns, including speed limits, right-of-way, stopping, yielding, following, and collision-avoidance criteria.Some methods use explicit labels or cost functions to rank socially and contextually plausible trajectories.
  • Reviewed approaches: Map-related knowledge forms the largest reviewed class of models, including lane-based, road-aware, and lane-aware prediction approaches.These methods use vectorized lane segments, curvilinear coordinates, lane boundaries, and feasible maneuvers as predictive information.

5. Uncertainty Estimation and Mitigation

The survey treats uncertainty as a central concern for robust trajectory prediction, distinguishing environmental uncertainty from system-internal uncertainty and aleatoric from epistemic uncertainty. It reviews uncertainty quantification and mitigation methods intended to support more cautious downstream planning.

  • Overview: Uncertainty-aware prediction is an active research area because uncertainty must be modeled across the autonomy stack for robust autonomous driving.The survey covers uncertainty sources, quantification techniques, and mitigation strategies.
  • Sources of uncertainty: External uncertainty includes hidden intentions, occlusions, adverse conditions, unexpected objects, rare events, and traffic-rule violations that expand plausible futures.Such factors can degrade prediction reliability, especially when rare events or violations occur.
  • Sources of uncertainty: Internal uncertainty reflects poor generalization, training-data bias, insufficient model capacity, and errors propagated from upstream perception and tracking modules.Upstream imprecision can affect downstream trajectory predictions in the detect-track-predict pipeline.
  • Uncertainty types: Aleatoric uncertainty arises from intrinsic randomness and sensing limitations, while epistemic uncertainty concerns uncertainty associated with the model and its knowledge.Examples include sensor noise, limited sensing range, calibration errors, low resolution, and multimodal pedestrian behavior.
  • Distributional uncertainty: Training-data representativeness is critical because distribution shifts in out-of-distribution traffic scenarios often reduce model performance.The survey highlights scenario coverage as a way to test whether training data captures relevant real-world variability.
  • Mitigation: Mitigation methods include physics-informed reconstruction, risk-aware attention, causal disentanglement, and end-to-end integration of detection, tracking, and forecasting.These approaches target missing inputs, underrepresented scenarios, spurious correlations, and cascaded uncertainty.
  • Quantification: Multimodal probabilistic forecasts quantify uncertainty by representing multiple plausible futures and estimating distributions or confidence measures over trajectories.Bayesian recurrent models, variational networks, and dropout-based analyses separate or combine aleatoric and epistemic uncertainty.

6. Planning-conditioned Trajectory Prediction

Planning-conditioned prediction uses planning objectives, behavioral assumptions, or anchors to forecast agents’ future trajectories. The survey compares model-based, learning-based, combined, goal-anchor, and trajectory-anchor strategies while identifying adaptability and scalability challenges.

  • Overview: Planning-conditioned methods assume that agents’ future trajectories reflect objectives, motion plans, or intended behavior rather than only past motion and scene context.This paradigm relies on observed or inferred plans and intentions to guide prediction.
  • Taxonomy: The survey categorizes planning-conditioned approaches as model-based, learning-based, combined, and anchor-based methods.The first three categories distinguish how planning is generated, while anchor-based methods use predefined or learned guidance points and paths.
  • Model-based planning: Model-based planning combines dynamics, behavioral models, control objectives, and road structure to generate interpretable and physically feasible trajectories.Its components include dynamics models, cost functions, and planners.
  • Learning-based planning: Learning-based planning uses data-driven models, often reinforcement learning, to infer strategic decisions, reactive control behavior, and multi-agent interactions.Inverse reinforcement learning can infer cost functions from observed pedestrian behavior and use them in a planner.
  • Goal anchors: Goal-anchor methods predict feasible destinations and condition trajectory generation on goals such as lane endpoints, intersections, or turn exits.This design aligns naturally with goal-driven tasks and may improve long-horizon forecasting through multiple goal hypotheses.
  • Trajectory anchors: Trajectory-anchor methods select or refine predefined or learned path templates, including straight, curved, and turning trajectories.PRIME combines planning-based search with learning-based prediction to generate reachable trajectory anchors.
  • Limitations: Model-based approaches offer interpretability and constraint satisfaction but lack adaptability, whereas learning-based methods handle complex behavior yet may ignore physical limits.Anchor-based methods provide semantic guidance but face scalability challenges in dense scenes.
  • Limitations: Efficient anchor generation remains a key challenge because scene analysis can be expensive and per-agent anchor generation may scale poorly as agent counts increase.The survey calls for compact, dynamically pruned anchor sets evaluated online rather than through pre-generated anchors.

7. Vision-Language Models for Prediction

This section reviews VLM-based trajectory prediction and related traffic-scene understanding models, along with language-enhanced datasets and integration challenges. VLM pipelines combine visual and textual inputs, but latency, prompt design, and output variability remain concerns for autonomous-driving deployment.

  • Scope of the review: The review covers a limited set of trajectory-prediction models and supplements them with traffic-scene understanding models for potential pipeline integration.It also surveys language-enhanced autonomous-driving datasets and challenges in integrating VLMs into the autonomous-driving stack.
  • Trajectory prediction models: Reviewed approaches use visual views, BEV representations, structured text, natural-language prompts, and language-model reasoning for difficulty ranking, lane-change prediction, and traffic-context understanding.Examples include two-stage VLM evaluation, text-enhanced CoverNet, LC-LLM, and LLM-based global traffic-context modeling.
  • Datasets: Language-enhanced datasets support tasks including object referring, action explanation, multi-object tracking, spatial understanding, and question answering in driving scenes.The surveyed datasets are built from sources such as nuScenes and BDD100K and add language-based annotations.
  • VLM foundations: VLM-based prediction models process visual and textual inputs through encoders and cross-modal attention to form task-specific representations.The general pipeline includes visual and textual encoders, joint representation learning, and a context-understanding layer.
  • Challenges and limitations: VLMs may provide common knowledge and contextual understanding for nuanced scene interpretation, but their integration into autonomous-driving systems still requires further study.The review identifies inference cost, prompt engineering, and non-deterministic outputs as practical challenges affecting deployment.

8. Collaboration between AVs for Prediction

Vehicle collaboration can compensate for restricted sensing range, occlusions, and sensor degradation by sharing complementary information. The section organizes collaborative approaches by early, intermediate, and late fusion while emphasizing bandwidth, latency, reliability, and deployment challenges.

  • Motivation: Vehicle collaboration can construct more comprehensive scene representations from complementary viewpoints, helping address restricted range, occlusions, and sensor degradation.Collaborating vehicles may share observations about vulnerable traffic agents and refine predictions using different viewing angles.
  • Efficient collaboration: Communication-reduction methods select complementary information or compress messages while fusion architectures address heterogeneous sensors, models, and alignment requirements.Examples include selective sharing, codebook-based compression, information filling, feature resizing, and hybrid fusion.
  • Fusion strategies: Collaborative systems are categorized by fusion stage: early fusion combines raw data, intermediate fusion shares semantic features, and late fusion aggregates independent decisions.The three modes impose different information-loss, bandwidth, communication, and noise trade-offs.
  • Latency and alignment: Transmission latency complicates temporal alignment because delays compound localization and transformation inaccuracies during collaborative perception.Latency-aware methods use synchronization and temporal-adjustment strategies to align exchanged information.
  • Open challenges: Task-specific collaboration remains underexplored, while practical systems must handle time sensitivity, inference-time limits, unreliable inputs, and verification of shared information.The review also calls for standardized testing platforms covering communication robustness, V2I collaboration, and diverse scenarios.

9. Discussion

The discussion finds that trajectory prediction has evolved into a broad family of hybrid and deep-learning methods, but accuracy gains are diminishing and deployment gaps remain. Emerging directions emphasize robustness, generalization, uncertainty, planning awareness, multimodal context, and systematic evaluation.

  • Existing solutions: Existing trajectory-prediction solutions span physics-based methods, classical machine learning, recurrent networks, generative models, graph architectures, attention mechanisms, and hybrid approaches.Current systems commonly combine CNNs, attention, encoder–decoder structures, graph-enhanced Transformers, and physics-based modeling.
  • Is the problem solved?: On the Waymo Motion Forecasting dataset, the top two solutions differ by only 0.01 minADE: MTR_v3: 0.55 versus ModeSeq: 0.56.The review uses this narrow leaderboard gap to illustrate diminishing accuracy improvements among strong models.
  • Is the problem solved?: Deployment readiness remains limited because averaged metrics obscure edge-case failures, cross-dataset evaluations are scarce, and instability at particular timesteps may require consistency mechanisms.These gaps make it difficult to assess generalization to underrepresented patterns and scenarios.
  • Emerging trends: Emerging research targets uncertainty reduction, collaborative perception, planning-conditioned prediction, VLMs, robustness, and generalization in dynamic environments.Planning-conditioned methods model planned trajectories and destinations to produce interpretable, constraint-compliant predictions in interactive scenarios.
  • Emerging trends: VLM deployment remains constrained by latency and predictability, especially when large models are hosted in the cloud or require substantial onboard computation.The review describes real-world VLM configuration and effectiveness as ongoing research questions.
  • Research gaps and future directions: Future work requires hardware-aware model selection, standardized cross-dataset benchmarking, noisy-perception robustness, and multi-horizon prediction.The review also identifies the need for unified scenario parametrization and evaluation of short- and long-horizon forecasting.

10. Conclusion

The review organizes autonomous-driving trajectory prediction methods and the broader prediction pipeline, then surveys emerging challenges and directions. It concludes that trajectory prediction remains open despite major advances, with hybrid domain-knowledge and data-driven methods offering a path toward balancing interpretability and predictive capacity.

  • Review scope: The review proposes a taxonomy distinguishing learning-based methods from non-learning methods, with learning-based approaches divided into machine learning, deep learning, and reinforcement learning.It also covers input and output modalities, benchmark datasets, and performance metrics used to assess prediction models.
  • Research gaps: The review emphasizes robustness and generalizability as continuing challenges requiring further research.These themes are presented alongside the study’s discussion of pressing challenges and evolving trends.
  • Research directions: The study examines uncertainty mitigation, collaborative perception, planning-conditioned prediction, goal-anchor methods, and VLM integration as active research directions.These directions address uncertainty, perception limitations, rational behavior modeling, interpretability, and contextual understanding.
  • Research gaps: VLM-based prediction remains subject to unresolved latency and scalability challenges.The review also identifies a need for systems that generalize across datasets and scenarios, supported by cross-dataset and qualitative evaluations.
  • Conclusion: Trajectory prediction remains an open and challenging research area despite significant advances over the past two decades.The field has progressed from physics-based models to Graph Transformers, while hybrid approaches combine domain knowledge with data-driven methods.
Loading 2503.03262v3…