Map-Free Trajectory Prediction

Updated 9 December 2025

Map-free trajectory prediction algorithms are techniques that forecast future agent trajectories without relying on fixed HD maps, using sensor data and historical context.
They employ dynamic graph encoding, multi-modal sensor fusion, and hierarchical decoding to capture spatial-temporal interactions and enforce kinematic constraints.
These approaches address limitations of map-based systems by handling outdated maps and reducing dependency on expensive spatial priors in dynamic environments.

Map-free trajectory prediction algorithms produce future motion forecasts for agents (typically vehicles or pedestrians) without the use of high-definition (HD) vector maps at inference, relying solely on sensor data, historical trajectories, or derived dynamic context. These approaches are designed to overcome intrinsic limitations of map-dependent systems, including map availability, cost, and susceptibility to outdated spatial priors, while aiming to maintain or surpass the performance of map-based counterparts in terms of accuracy, efficiency, and robustness.

1. Core Principles and Problem Formulation

Map-free trajectory prediction models eschew explicit road topology during deployment. The canonical task is to predict $K$ plausible future trajectories $Y \in \mathbb{R}^{K \times T_f \times 2}$ for each agent given its observed trajectory history $X$ , possibly augmented with sensor inputs (e.g. camera, LiDAR). The absence of map priors at inference introduces challenges: loss of geometric road constraints and semantic lane association, necessitating compensatory mechanisms for behavioral and contextual reasoning (Liu et al., 17 Nov 2024, Liao et al., 2 May 2024, Zhang et al., 24 Jul 2025).

Recent paradigms formalize the process as either direct sequence modeling, dynamic context graph encoding, or through implicit scene representation in sensor space (e.g. BEV feature maps) (Kong et al., 12 Sep 2025, Xiong et al., 2 Dec 2025). Collision avoidance and kinematic feasibility must be enforced via architectural inductive biases, auxiliary losses, or by distilling map knowledge during training (Wang et al., 2023, Liu et al., 17 Nov 2024).

2. Algorithmic Mechanisms and Architectural Innovations

Dynamic Spatio-Temporal Encoding

Rich dynamic context is encoded by various mechanisms:

Adaptive Structural Graphs: Agent-centric dynamic graphs are constructed solely from observed positions with connections defined by proximity or kinematic relevance (e.g., radius-based adjacency in (Liao et al., 2 May 2024)).
Agent Interaction Blocks: Stacked GCNs or attention layers capture spatial (relative position, heading) and temporal dependencies, typically via multi-head self-attention, positional encodings, and adaptive edge attribute learning (Xiang et al., 2023, Liao et al., 2 May 2024, Liu et al., 17 Nov 2024).
Sensor-Level Feature Extraction: Raw sensor inputs (images, LiDAR) are fused into BEV representations without explicit map priors, with deformable attention mechanisms allowing flexible, data-driven contextual aggregation (Kong et al., 12 Sep 2025).

Behavioral and Frequency Domain Modules

Agent behavior is modeled via centrality-based metrics, VRNN encoders, or direct frequency-domain processing:

Behavior-Aware Embeddings: Node-level embeddings summarize degree, closeness, eigencentrality, betweenness, power, and Katz centrality (plus their temporal derivatives) over the graph, followed by VRNN+GRU encoders (Liao et al., 2 May 2024).
Frequency-Selective Information: Mixture-of-Experts in the frequency domain combined with selective time/patch attention boost robustness to aliasing and redundant signal contaminants in historical data (Xiong et al., 2 Dec 2025).

Hierarchical Aggregation and Decoding

Multiple models use hierarchical aggregation (e.g., multi-scale cross-attention, hierarchical query fusion) and iterative decoding:

Hierarchical Feature Aggregation: Aggregation over multiple temporal scales (sampling every $2^{h-1}$ timestep for hierarchy level $h$ ) allows multi-scale trajectory query formation (Liu et al., 17 Nov 2024).
Recursive and Iterative Decoding: Principle of recursively refining coarse predictions (coarse-to-fine or global-to-local) to limit error propagation and guarantee smoothness. For example, G2LTraj first predicts global keysteps, then recursively fills intermediate points, enforcing spatial and temporal constraints at each scale (Zhang et al., 30 Apr 2024).

Uncertainty Modeling and Scenario Gating

Certain methods explicitly estimate the uncertainty inherent in online-generated scene information and learn to adaptively fuse this uncertainty depending on ego-vehicle predicted kinematics:

Covariance-Based Map Uncertainty: Geometry-aware uncertainty (full $2 \times 2$ Gaussian covariance $\Sigma$ ) regressed per BEV map vertex (Zhang et al., 24 Jul 2025).
Proprioceptive Scenario Gating: Learned gating chooses whether to use uncertainty-augmented prediction based on the agent's forecasted yaw-rate, directly aligning model confidence with driving scenario (Zhang et al., 24 Jul 2025).

3. Map Knowledge Distillation and Hybrid Training Paradigms

Several SOTA map-free approaches exploit map-based training signals via knowledge distillation, allowing student (map-free) models to inherit scene compliance and multi-modal reasoning from map-privileged teachers (Wang et al., 2023, Liu et al., 17 Nov 2024):

Feature Distillation: Student features are matched via variational L2 losses to teacher internal representations, including those downstream of map-encoder branches.
Output Distillation: Student multimodal outputs (e.g., mixture parameters) are aligned with teacher distributions using cross-entropy and regression losses, forcing the student to recover topological-awareness absent during inference.
Intermediate Query Matching: MFTP (Liu et al., 17 Nov 2024) matches hierarchical encoder and intermediate decoder queries between teacher and student, combined with standard regression and classification losses.

The teacher–student framework ensures that, although deployment is map-free, performance and path topology closely approach map-based upper bounds (Wang et al., 2023, Liu et al., 17 Nov 2024).

4. Notable Model Frameworks

The table summarizes representative map-free trajectory prediction algorithms, highlighting core mechanisms and distinguishing features.

Model	Key Mechanism	Map Priors (Inference)	Training Approach
G2LTraj (Zhang et al., 30 Apr 2024)	Global-to-local, spatial & temporal constraints, granularity selection	None	Plug-in head for any predictor
Knowledge Distillation (Wang et al., 2023), MFTP (Liu et al., 17 Nov 2024)	Teacher–student, distillation Loss	None	Map-based teacher, map-free student
MFTraj (Liao et al., 2 May 2024)	Behavior/centrality, SAIGCN, Linformer	None	Single/graph encoding, behavior-aware
BEVTraj (Kong et al., 12 Sep 2025)	BEV sensor fusion, deformable attention, sparse goal proposals	None	End-to-end in BEV space
MoE+Selective Attention (Xiong et al., 2 Dec 2025)	Frequency-domain MoE, temporal/spatial selective attention	None	Multi-domain, multi-scale
Fast Model (Xiang et al., 2023)	Two-stage agent/interaction encoder, LSTM, GCN, transformer	None	Parallel spatial/temporal pipelines
Mapping Uncertainty (Zhang et al., 24 Jul 2025)	Online map gen., scenario gating, covariance fusion	Online sensor map only	Kinematic-aware, uncertainty fusion
EKF+ML+CurveFit (Agrawal et al., 2020)	EKF filtering, shape recognition, parametric propagation	None	Explicit geometric class fit

5. Quantitative Evaluation and Real-World Performance

On public benchmarks such as Argoverse, ETH/UCY, nuScenes, MoCAD, HighD, and NGSIM, leading map-free models now match or marginally lag behind SOTA map-based approaches in minADE, minFDE, and Miss Rate metrics:

G2LTraj delivers up to 13.3% lower FDE and up to 12.1% lower ADE than simultaneous baselines, and outperforms recursive methods on ETH/UCY (Zhang et al., 30 Apr 2024).
MFTraj achieves minADE=1.59 m, minFDE=3.51 m, outperforming most map-based SOTA and remaining robust to up to 50% missing observations (Liao et al., 2 May 2024).
BEVTraj yields minADE $_{10}$ =0.94 m and minFDE $_{10}$ =2.05 m on nuScenes (50 m range), comparable to HD map-based MTR and Wayformer but with lower Miss Rate (Kong et al., 12 Sep 2025).
Knowledge-distilled map-free models recover 85–95% of map-dependent performance, with consistent 3–14% gains in ADE/FDE over standard map-free baselines (Wang et al., 2023, Liu et al., 17 Nov 2024).
Mapping Uncertainty methods demonstrate up to 23.6% Miss Rate reduction through adaptive covariance fusion and scenario gating (Zhang et al., 24 Jul 2025).

These results are enabled by incorporating strong agent–agent context, carefully crafted multi-scale or scenario-dependent features, and explicit or implicit transfer of topological knowledge.

6. Limitations, Open Challenges, and Practical Considerations

While modern map-free algorithms close the performance gap with map-based systems, several caveats remain:

Lack of road-rule compliance: Without semantic maps, prediction compliance with traffic norms (e.g., stoplines, right-of-way) depends on learned priors, which may fail in novel topologies (Kong et al., 12 Sep 2025, Xiong et al., 2 Dec 2025).
Generalization: In situations with long-range occlusion, poor sensor coverage, or sparse contextual cues, map-free models may underperform, particularly in rare or out-of-distribution configurations.
Efficiency vs. Interaction Complexity: Fully connected agent graphs scale as $O(N^2)$ ; although efficient approximations (Linformer, selective attention, sparsification) exist, scalability is an active area of research (Liao et al., 2 May 2024, Xiang et al., 2023).
Dynamic Scenario Adaptation: Online adaptation to changing road geometry is a strong point (especially in BEV-based models), but may come at the expense of explicit topology and interpretability (Kong et al., 12 Sep 2025, Zhang et al., 24 Jul 2025).

A plausible implication is that integrating scenario-aware uncertainty fusion and hierarchical/multi-modal representations yields the best trade-offs for deployment in dynamic, map-starved environments.

7. Future Directions and Research Outlook

Current trends suggest continued convergence of multi-modal sensor fusion, knowledge distillation from map-based supervisors, and scenario-adaptive uncertainty integration:

Unified end-to-end models operating directly on raw sensor data (images, LiDAR), leveraging BEV or graph-based representation, are anticipated to dominate.
Distillation and transfer learning: Robust approaches to transferring topological priors, traffic norms, and lane association from privileged teachers to lightweight deployable students remain a key area (Liu et al., 17 Nov 2024, Wang et al., 2023).
Modular scenario gating and explainability: Learning when to trust (or bypass) various modalities, calibrated by agent kinematics or local environment, will be critical for safety and interpretability (Zhang et al., 24 Jul 2025).
Beyond driving: The principles in map-free trajectory prediction—multimodal context modeling, coarse-to-fine recursion, selective attention—are likely applicable to broader domains including aerial robotics, pedestrian simulation, and smart-city infrastructure.

Thus, map-free trajectory prediction algorithms offer an increasingly viable alternative to HD map-dependent pipelines, combining architectural, training, and inference innovations to achieve robust, context-aware, and scalable motion forecasting in real-world settings.