PaperlayerView exact revision on arXiv

Source-linked AI summary

Trends in Motion Prediction Toward Deployable and Generalizable Autonomy: A Revisit and Perspectives

Letian Wang, Marc-Antoine Lavoie, Sandro Papais, Barza Nisar, Yuxiao Chen, Wenhao Ding, Boris Ivanovic, Hao Shao, Abulikemu Abuduweili, Evan Cook, Yang Zhou, Peter Karkus, Jiachen Li, Changliu Liu, Marco Pavone, Steven Waslander

arXiv:2505.09074v4cs.RO

TL;DR

Motion prediction research faces a gap between deployment needs and benchmark settings, as well as limited data and open-world shifts. This survey reviews methods and practices for deployability and generalization; reviewed studies report gains from larger, more diverse data and scaling, while highlighting limits in evaluation and data coverage.

  • Problem

    Motion prediction benchmarks and datasets do not fully cover closed-loop deployment needs or the broad, shifting conditions of open-world settings.

    Source: §3 · pp. 42–43, §4.7 · pp. 117, 122

  • Method

    The survey synthesizes motion-prediction methods and perspectives on integrating prediction into autonomy stacks and extending generalization to open-world settings.

    Source: §3 · p. 42, §5 · p. 123

  • Results

    Reviewed studies report stronger generalization from larger, more diverse training data, consistent cross-dataset gains from SmartPretrain, and power-law performance scaling with compute using 500,000 hours of driving data.

    Source: §4.6 · pp. 115–116, §4.6.1 · p. 117

  • Takeaways & Limitations

    Closed-loop metrics improve with scaling and are described as more useful than open-loop metrics for selecting models for downstream deployment.

    Source: §4.6.1 · p. 117

    Existing curated benchmarks isolate prediction from raw sensing and downstream planning, while end-to-end benchmarks can assess driving performance at the expense of evaluating prediction.

    Source: §3 · p. 43

Abstract

from arXiv · show

Motion prediction, recently popularized as world models, refers to the anticipation of future agent states or scene evolution, which is rooted in human cognition, bridging perception and decision-making. It enables intelligent systems, such as robots and self-driving cars, to act safely in dynamic, human-involved environments, and informs broader time-series reasoning challenges. With advances in methods, representations, and datasets, the field has seen rapid progress, reflected in quickly evolving benchmark results. Yet, when state-of-the-art methods are deployed in the real world, they often struggle to generalize to open-world conditions and fall short of deployment standards. This reveals a gap between research benchmarks, which are often idealized or ill-posed, and real-world complexity. To address this gap, this survey revisits the generalization and deployability of motion prediction models, with an emphasis on applications of robotics, autonomous driving, and human motion. We first offer a comprehensive taxonomy of motion prediction methods, covering representations, modeling strategies, application domains, and evaluation protocols. We then study two key challenges: (1) how to push motion prediction models to be deployable to realistic deployment standards, where motion prediction does not act in a vacuum, but functions as one module of closed-loop autonomy stacks - it takes input localization and perception, and informs downstream planning and control. 2) How to generalize motion prediction models from limited seen scenarios/datasets to the open-world settings. Throughout the paper, we highlight critical open challenges to guide future work, aiming to recalibrate the community's efforts, fostering progress that is not only measurable but also meaningful for real-world applications. The project webpage can be found here https://trends-in-motion-prediction-2025.github.io/.

1. Introduction

Motion prediction must bridge perception and downstream action despite behavioral uncertainty, noisy inputs, real-time constraints, and distribution shifts. This survey frames deployability and open-world generalization as connected challenges, then organizes relevant methods and evaluation practices across autonomous driving, robotics, and related fields.

1.1 Background

  • Closed-loop deployment remains poorly reflected by benchmarks that assume clean inputs, isolate prediction, and measure open-loop accuracy rather than whole-system performance.The survey identifies interface design, uncertainty propagation, cross-module learning, and system-level evaluation as challenges for deployable prediction.Source: §1.1 · p. 6
  • Finite datasets and domain-specific benchmarks leave open-world generalization underexplored, although real deployments encounter unseen agents, behaviors, contexts, and operating conditions.Optimizing a fixed-domain metric can come at the cost of generalization and degrade performance in unseen settings.Source: §1.1 · pp. 6–7
  • The proposed generalization lifecycle links diverse data collection, robust training, safety-aware deployment, and test-time adaptation in a recurring cycle.Detected out-of-distribution data can be selectively stored and annotated to expand training coverage for the next cycle.Source: §1.1 · p. 7

1.3 Paper’s Scope and Framework

  • The survey spans representations, modeling approaches, applications, and evaluation, but reports that robotics prediction lags driving amid less standardized datasets and protocols.Driving benefits from large-scale public benchmarks that have helped standardize evaluation practices.Source: §1.1 · p. 7, §1.3 · pp. 9–10

1.2 Problem Formulation

  • The task maps observations from a recent time window to predicted future agent states or scene evolution, with representations varying by task and sensor setting.Outputs may be deterministic or probabilistic to account for multiple plausible futures.Source: §1.2 · pp. 7–8
  • Deployability calls for compact, task-relevant inputs, planner-compatible outputs, uncertainty handling, cross-module training, and evaluation in the full autonomy stack.The evaluation target is ultimately closed-loop system performance.Source: §1.2 · p. 8

1.3 Paper’s Scope and Framework

  • The survey centers on learning-based motion prediction and reviews deployment and generalization, while also covering related world models, motion generation, and learned planning.These areas often share architectures and learning paradigms, and the survey says many insights apply across them.Source: §1.3 · pp. 8–9

2. Taxonomy of Motion Prediction Methods

The section provides a supporting taxonomy of motion prediction methods, organized around representations, modeling approaches, target applications, and evaluation strategies.

2. Taxonomy of Motion Prediction Methods

  • The taxonomy organizes motion prediction methods by representation, modeling approach, target application, and evaluation strategy, including commonly used metrics and benchmark datasets.It serves as background for the paper’s main focus on deployability and generalizability.Source: §2 · p. 14

2.1 Motion Factors and Common Representations

Motion behavior reflects internal agent factors, interactions with surrounding agents, and environmental context. The section surveys representations that encode these factors as inputs and describe predicted behavior as outputs.

2.1.1 Motion Factors

  • Historic motion and pose, intent and ongoing task, and semantic attributes such as agent type and individual preferences are internal factors shaping an agent’s behavior.Source: §2.1.1 · p. 14
  • Surrounding agents influence motion through their histories, intentions, and classes, as well as through the target agent’s choices to ignore, interact, negotiate, or coordinate.Source: §2.1.1 · p. 14
  • Environmental context ranges from open space to obstacles, scene geometry and topology, and semantics such as traffic lights, lane markings, and driving customs.Source: §2.1.1 · p. 14

2.1.2 Input and Output Representation

  • Common agent-centric representations include waypoint trajectories, articulated poses, and intentions or tasks, which can serve as inputs, outputs, or intermediate prediction targets.Trajectories encode states over time; articulated poses add body-part and joint detail, while intentions can include goals, skills, navigation signals, or tasks.Source: §2.1.2 · pp. 14–15, Figure 6 · p. 15
  • Environment maps range from landmarks and scene graphs to rasterized and vectorized maps; occupancy grids represent agent presence, while occupancy flow adds motion vectors.Vectorized maps aim to avoid lossy rendering and capture topology; flow vectors indicate motion direction and magnitude in each grid cell.Source: §2.1.2 · pp. 15–16, Figure 6 · p. 15
  • Models may take raw images or point clouds as input and predict in the same format, or use upstream perception features as a mid-level alternative to raw signals and trajectories.Direct sensor-input approaches bypass object detection and tracking; upstream features include detection embeddings, tracking affinity maps, online map features, and visual tokens.Source: §2.1.2 · p. 16
  • Inputs are commonly deterministic, but can represent upstream uncertainty probabilistically; prediction outputs are typically probabilistic to reflect human behavior’s inherent uncertainty and multiple possible futures.Probabilistic outputs support uncertainty quantification and calibration for downstream decision-making.Source: §2.1.2 · pp. 16–17

2.2 Modelling Approaches

This section surveys motion-prediction modeling approaches, organizing methods into physics-based, planning-based, and deep-learning-based categories while noting that categories may be combined. It also describes interaction modeling with attention and approaches for representing multimodal future motion and uncertainty.

2.2.1 Physics-Based Methods

  • Physics-based methods propagate hand-crafted motion models or rules, using either one model under fixed assumptions or several models for adaptive prediction.The models are based on Newton’s laws; single-model and multi-model approaches are the survey’s two broad subdivisions.Source: §2.2.1 · p. 17

2.2.1.1. Single-model approaches

  • Kinematic models simplify motion by ignoring forces, while dynamic models account for forces but may require more parameters, force estimation, and computation.Uncertainty can be represented with Gaussian mixtures or Monte Carlo sampling and updated using Kalman filtering; for trajectory prediction, models that abstract away forces are often preferred.Source: §2.2.1.1 · pp. 17–18

2.2.1.2. Multi-model approaches

  • Multi-model methods combine motion patterns or maneuvers and estimate their probabilities to switch among models or fuse their predictions.Model sets may be fixed or adapted online, and range from straight-or-turning primitives to semantically defined maneuvers such as overtaking and merging.Source: §2.2.1.2 · p. 19

2.2.2 Planning-Based Methods

  • Planning-based methods predict motion by assuming agents optimize an objective, either by planning with a known objective or inferring it from observations.The survey distinguishes forward planning from inverse planning, which learns the objective function before predicting motion.Source: §2.2.2 · p. 19

2.2.2.1. Forward Planning Methods

  • Forward planning uses objectives for task completion, safety, smoothness, and sometimes control effort, solved through optimization, search, or sampling.These methods can incorporate dynamic constraints to ensure physical realizability; probabilistic and interactive variants model uncertainty, multiple goals, or other agents’ responses.Source: §2.2.2.1 · p. 20

2.2.2.2. Inverse Planning Methods

  • Inverse planning learns an agent’s objective from observed expert behavior under a rationality assumption, commonly through inverse reinforcement learning.Learned rewards can capture driving styles, hierarchical intentions, or scenario-specific objectives, enabling predictions to adapt to individual preferences and contexts.Source: §2.2.2.2 · p. 21

2.2.3 Deep-Learning-Based Methods

  • Physics- and planning-based methods suit predefined layouts and routine behavior but struggle with novel, complex settings; deep learning learns patterns from data and dominates large-scale benchmarks.The survey attributes traditional methods’ limits to restricted expressiveness and task-specific rule design; neural networks learn relationships among agents and environmental factors, while mismatched inductive biases can impair performance [46].Source: §2.2.3 · p. 21
  • RNNs encode temporal histories, CNNs model spatial structure on grids, GNNs aggregate graph-neighbor relationships, and attention weights influences among features.RNNs support trajectory or intention prediction and interaction encoding; CNNs process agent and map grids but face receptive-field and computation trade-offs, while vectorized-map graphs support agent-map reasoning. Temporal attention is parallelizable, unlike sequential RNN calculations.Source: §2.2.3 · pp. 22–25
  • Agent attention aggregates neighboring agents’ effects on a target, and some methods constrain attention to graph-connected or unmasked agents [561].Attention weights capture each neighbor’s effect on the target at different levels.Source: §2.2.3 · p. 25
  • Agent-map attention treats map elements as entities and combines map self-attention with cross-attention to integrate spatial relationships and semantic cues into agent motion predictions.Temporal, agent, and agent-map attention are often combined to reason about scene information across spatial and temporal dimensions.Source: §2.2.3 · pp. 25–26
  • Space-based attention represents scenes as spatial entities, using grid cells or dynamic insertion areas (DIAs) to model agent-map relationships and potential movement gaps.DIA attention models relationships among candidate areas formed by agents and map markings; related concepts include lane homotopies and topological braids.Source: §2.2.3 · p. 26
  • Anchors, probabilistic models, and generative models offer complementary ways to represent diverse futures, respectively structuring prediction, estimating probabilities, and sampling complex multimodal distributions.Generative approaches include CVAEs, GANs, and diffusion models; their diversity also supports motion generation for simulation and data generation.Source: §2.2.3 · pp. 26–27

2.3 Application Domains

Motion prediction spans domains with overlapping methods and domain-specific requirements. The section describes how applications shape prediction targets, inputs, and planning needs.

2.3.1 Autonomous Driving and Navigation

  • Autonomous driving predicts short-term ground-level states of road users to support conflict-free planning, using trajectories, intentions, occupancy, or flow representations and physics-, planning-, or learning-based methods.Models may incorporate agent histories, road geometry, and traffic signals and regulations.Source: §2.3.1 · pp. 27–28

2.3.2 Human-Robot Interaction

  • HRI prediction supports safe collaboration across manufacturing, warehouse, home, medical, and cooking robots, and can target 2D agent states or articulated 3D poses.Inputs may include agent histories, scene information, and tasks or intentions.Source: §2.3.2 · pp. 28–29, Figure 12 · p. 29
  • In human-robot interaction, local collision avoidance favors fast short-term forecasts, while safety-critical applications can use reachability methods to guarantee local collision avoidance.Linear velocity projections may suffice in the simplest cases.Source: §2.3.2 · pp. 29–30
  • Global HRI planning favors long-term, multi-hypothesis or intention predictions that account for contextual cues and the uncertainty of human behavior.Compared with local planning, real-time reactivity is usually relaxed.Source: §2.3.2 · p. 30

2.3.3 Visual Motion Understanding

  • Visual motion understanding uses prediction for surveillance, monitoring, crowd-flow analysis, interactive games, and VR/AR experiences.Surveillance can use long-term forecasts for track association and crowd analysis; games use prediction to respond to player actions and resolve tracking ambiguities.Source: §2.3.3 · pp. 30–31, Figure 13 · p. 30

2.3.4 General Time-Series Prediction

  • Motion prediction is a specialized form of time-series prediction in which the temporal dynamics describe physical agent motion in space.Related methods also apply to forecasting in fields such as healthcare, energy, finance, and transportation.Source: §2.3.4 · p. 31

2.4 Evaluation Strategy

This section surveys evaluation metrics tied to prediction representations and datasets across driving and human-motion tasks. It also addresses dataset diversity and inconsistent splits that complicate result reuse and fair comparison, alongside multi-dataset benchmarks and unified tools for cross-dataset evaluation.

2.4.1 Performance Metrics

  • Trajectory metrics cover accuracy, coverage, safety, and uncertainty, with variants accounting for multimodal predictions and confidence.Accuracy metrics assess endpoints or whole trajectories; coverage metrics measure predictions within acceptable ranges or drivable areas, while safety and uncertainty metrics capture other aspects of prediction quality.Source: §2.4.1 · pp. 32, 34, 35
  • mAP filters predictions by confidence thresholds and computes precision-recall area, while soft mAP does not penalize additional matching predictions.Standard mAP allows only the highest-confidence match as a true positive; soft mAP ignores other matches, making it more lenient for multi-hypothesis prediction.Source: §2.4.1 · p. 35
  • Existing metrics do not directly measure predictive uncertainty sufficiently, motivating distribution-based measures such as KL divergence and likelihood.Negative log likelihood and predicted probability are cited as more commonly used variants when directly measuring the ground-truth trajectory distribution is challenging.Source: §2.4.1 · pp. 35–36
  • Metrics are representation-specific: intention tasks use classification accuracy, occupancy uses IoU-family measures and mAP, and flow uses endpoint error.Occupancy measures include AUC and Soft-IoU; endpoint error measures the Euclidean distance between predicted and ground-truth flow vectors.Source: §2.4.1 · pp. 36–37

2.4.2 Datasets

  • Driving and human-motion datasets use vehicle or drone cameras, cameras or motion capture, and encompass varied scene types and data modalities.Driving datasets may provide standardized splits and horizons, trajectories, maps, and auxiliary signals; Waymo also includes occupancy and flow tasks.Source: §2.4.2 · pp. 37–38
  • Human-motion datasets span urban, campus, indoor, shopping-center, and simulated scenes, with varied sensors and specialized human-robot task datasets [6].Collection conditions range from CMU Mocap at 120 Hz to ATC recordings spanning 92 days; TH¨OR combines cameras, LiDAR, and motion capture.Source: §2.4.2 · pp. 38–40
  • Official horizons are available in some datasets, but training, validation, and test splits are often customized, hindering result reuse and fair comparison.UCY, ETH, and ForkingPaths provide historic and future horizons, while official splits are usually absent.Source: §2.4.2 · pp. 39–40
  • Datasets represent motion as 2D trajectories, 3D poses, or body meshes, and may add maps, gaze, intention, behavior, bounding boxes, or video.These formats and auxiliary signals support goals from trajectory forecasting to fine-grained pose estimation and motion analysis with visual data.Source: §2.4.2 · p. 40
  • Multi-dataset resources support broader evaluation: Trajnet and Trajnet++ combine five sources, OpenTraj covers more than 20 datasets, and Trajdata provides a unified interface [482, 294, 21, 250].ScenarioNet aggregates real-world datasets for planning development and evaluation, while UniTraj and SmarPretrain unify driving-data formats for large-scale training and cross-dataset evaluation [335] [153] [679].Source: §2.4.2 · pp. 40–41

3. Deploying Motion Prediction within Real-World Autonomous Systems

This section examines how motion prediction moves from isolated benchmark settings into real-world autonomy stacks, where it must integrate with perception and planning/control under uncertainty and feedback. It covers representations, uncertainty handling, and joint learning across modules, along with trade-offs in computation, interpretability, temporal consistency, and real-time modeling.

3. Deploying Motion Prediction within Real-World Autonomous Systems

  • Deployment demands efficient stack-compatible representations, uncertainty/error propagation, joint evaluation across modules, and closed-loop tests of feedback, latency, consistency, safety, and task performance.Current standalone benchmarks omit downstream planning, while end-to-end benchmarks may not evaluate motion prediction itself; a unified open-source protocol remains elusive.Source: §3 · pp. 42–43

3.1.1 Introduction to the Autonomy Stack

  • Autonomy stacks map sensor inputs through perception, prediction, and action; object trajectories provide compact interfaces, while occupancy grids support scene-level forecasting and planning.Inputs can include cameras, LiDAR, radar, event cameras, and HD maps; combining sensors can improve robustness but raises computation, cost, and fusion challenges.Source: §3.1.1 · pp. 43–44

3.1.2 Autonomy Tasks Decomposition

  • Modular pipelines aid interpretability but bottleneck information and misalign objectives; monolithic end-to-end models remove interfaces but challenge safety, while modular end-to-end designs retain explicit tasks.Modular end-to-end designs explore sequential, parallel, and hybrid connections between modules.Source: §3.1.2 · pp. 45–46

3.1.3 Motion Prediction in End-to-End Modular Methods

  • Joint perception-prediction and prediction-planning methods exchange information across modules, condition forecasts on ego plans, and jointly optimize decisions; multi-task complexity and reduced interoperability remain.Ego-conditioned prediction can be used in sampling-based or optimization-based plan selection, and prediction and planning can be jointly trained.Source: §3.1.3 · pp. 46–48

3.2 Perception-Prediction Integration

  • Perception-prediction integration requires compatible representations and interfaces, calibrated propagation of upstream uncertainty and errors, and joint learning with temporal information exchange.These design choices aim to improve information flow, computation, robustness, and alignment with overall system performance.Source: §3.2 · p. 48

3.2.1 Representations

  • Object-based representations preserve identity and interpretability but accumulate detection/tracking errors; occupancy captures fine irregular shapes but loses identity and raises computation costs.Latent representations can support modularity and real-time efficiency, yet compression and non-differentiable operations may discard contextual information.Source: §3.2.1 · pp. 49–50
  • Latent representations enable continuous gradient and uncertainty flow for joint learning, but can reduce interpretability, require more data and computation, and complicate module interoperability.Fine-grained upstream context can support more informed downstream decisions, while replacing a component may require retraining for compatibility.Source: §3.2.1 · p. 50
  • Tracking-free methods avoid hard identity assignments and tracking errors by predicting from detections, but maintaining long-term temporal consistency without explicit identity links remains a challenge.Soft associations aggregate detections over time, whereas detection-based methods predict directly from per-frame detections and can aggregate historical features.Source: §3.2.1 · p. 51

3.2.2 Dealing with Perception Uncertainty

  • Deterministic perception outputs can lose ambiguity and pass detection, tracking, or mapping errors into prediction, making uncertainty propagation, calibration, and reduction important for reliable autonomy.Probabilistic uncertainty modeling must balance fidelity against the computational efficiency needed for real-time inference.Source: §3.2.2 · p. 52
  • Prediction can incorporate detection-state distributions and class probabilities, but most models overlook object-presence uncertainty, leaving false positives and missed detections an open robustness challenge.Passing rich perception feature maps is identified as a promising alternative to relying only on deterministic object presence.Source: §3.2.2 · p. 53
  • Probabilistic tracking approaches represent association ambiguity with multiple hypotheses or soft affinities and can propagate track-localization uncertainty into prediction.Tracking-free prediction instead avoids tracking-induced errors by predicting directly from detections, but sacrifices tracking information.Source: §3.2.2 · pp. 53–54
  • Online map uncertainty estimates address sources including occlusion, sensor range, and weather; incorporating map distributions improved prediction accuracy by up to 15% and accelerated training convergence by 50%.Occupancy methods estimate uncertainty for current scenes, while uncertainty in future occupancy forecasting remains underexplored.Source: §3.2.2 · p. 54

3.2.3 Joint Learning and Prediction-Informed Perception

  • Joint perception and prediction can reduce information bottlenecks and error accumulation, but explicit interfaces impede gradient flow through operations such as thresholding and hard association.Differentiable latent interfaces—including dense occupancy feature maps and sparse query tokens—support end-to-end learning with varying computational costs.Source: §3.2.3 · p. 55
  • Prediction-informed temporal perception feeds predicted motion back into later perception cycles, helping address occlusion, sensor limitations or degradation, and tracking errors.Temporal detection aggregates historical features; prediction-informed detection propagates motion to augment point clouds or initialize queries, while prediction-informed tracking uses short-term forecasts, including during brief occlusions.Source: §3.2.3 · pp. 55–56, Figure 18 · p. 55

3.3 Prediction-Planning Integration

This section examines prediction-planning integration, emphasizing representations, uncertainty handling, and outputs that support robust, efficient, behaviorally consistent closed-loop decisions. It covers planning-informed and joint approaches that account for interactions between ego actions and other agents, along with their potential benefits and computational, data, and safety challenges.

3.3.1 Representations

  • Object-, occupancy-, and latent representations trade interpretability and planning compatibility against open-world expressiveness, computation, and safety; alternatives include MATS [246], topology patterns [572], and sensor-level video [225] and point clouds [278].Object-based outputs are compact and interpretable but can lose information and struggle with novel classes; occupancy captures arbitrary motion but is dense, while latent representations support joint optimization but are less interpretable. MATS align with system dynamics, and topology-aware patterns encode agent–map relations; action-conditioned video offers a planning signal but still faces geometric-consistency and evaluation challenges.Source: §3.3.1 · pp. 57–59

3.3.2 Deal with Prediction Uncertainty in Planning

  • Prediction uncertainty includes reducible epistemic ignorance, irreducible aleatoric variation, and sampling-induced temporal inconsistency; planners need diverse but calibrated plausible futures rather than collapsed or excessive modes.Aleatoric uncertainty includes uncertainty about agents’ intentions and how they execute a chosen movement. High-frequency latent sampling can make trajectories oscillate over time and reduce planning performance [97].Source: §3.3.2 · pp. 59–60
  • Calibration remains underexplored: motion predictors often report likelihoods or variances without testing whether they match observed outcomes, and may misstate uncertainty in familiar versus novel interactive scenes.The paper notes that benchmarks rarely reward calibration; conformal methods [631] [354] often target post-hoc coverage rather than calibration of the model’s own uncertainty outputs.Source: §3.3.2 · p. 61
  • Uncertainty reduction can use historical posteriors, active information gathering, ego-conditioned predictions [96], and joint scene-consistent predictions [96].Active actions can elicit information about other agents, while ego conditioning models how planned behavior affects them; joint prediction constrains futures to be coherent, feasible, and collision-free.Source: §3.3.2 · pp. 61–62

3.3.3 Planning-Informed Prediction and Joint Learning

  • Planning and prediction are bidirectional, but marginal predictions and agent-level benchmarks can miss globally inconsistent scenes that are less useful to downstream planners.Joint, scene-centric prediction [96] targets coherent multi-agent futures, though enumerating topological scene modes can become computationally costly as agents and map elements increase.Source: §3.3.3 · pp. 62–63
  • Ego-conditioned prediction focuses computation on planner-relevant agents or regions, while end-to-end joint training [233, 272, 239, 240] prioritizes agents through planning loss.Conditioning on hypothetical ego trajectories [11] [96] has been reported to improve prediction accuracy with ground-truth ego motion and downstream planning performance [97].Source: §3.3.3 · p. 63
  • Sampling-based policy planning conditions other-agent predictions on candidate ego trajectories; DTPP [239] prunes low-score tree nodes, nearly halving runtime without compromising planning performance.DTPP also uses different sampling granularities across stages and predicts over the whole ego trajectory tree to avoid redundant computations.Source: §3.3.3 · pp. 63–64
  • Optimization-based policy planning trades sampling costs for optimization overhead: IJP [99] jointly optimizes ego and agent motions across route-planned interaction modes, while MATS [246] optimizes ego motion through predicted agent dynamics.IJP begins with unconditioned predictions; MATS predicts parameters governing surrounding agents’ future dynamics rather than trajectories directly.Source: §3.3.3 · p. 64
  • Game theory [240] and cumulative prospect theory provide rational and reactive behavior biases for policy planning, but their assumptions may constrain expressiveness.Source: §3.3.3 · p. 64
  • Joint prediction and planning learns surrounding-agent futures and feasible ego trajectories together, offering data-driven human-like behavior but requiring diverse, rich training data [239] [240] [360].Unlike rule-based planning, it avoids extensive hand-engineering and brittle rule interactions, but lacks explicitly injected human knowledge.Source: §3.3.3 · pp. 64–65
  • Joint prediction-planning remains decoupled from perception, supporting inspection and debugging; hybrid systems reuse mature perception modules while focusing joint learning on data-scarce motion tasks.The separation can aid assessment of safety and robustness, especially in safety-critical settings.Source: §3.3.3 · p. 65
  • Key research challenges are acquiring interaction-rich training data, scaling tractable multi-agent reasoning, and integrating uncertainty estimation and safety constraints into learning objectives.These challenges are described as relatively underexplored for joint prediction and planning.Source: §3.3.3 · p. 65

3.4 Evaluation

The section focuses on evaluating motion prediction for deployment by accounting for the interactions among perception, prediction, planning, and control in autonomy systems. It also highlights efficiency and integration as considerations for assessing performance under realistic conditions.

3.4 Evaluation

  • Realistic motion-prediction evaluation should capture upstream perception, downstream planning, and closed-loop system effects rather than assess prediction in isolation.Existing benchmarks often use curated perception inputs and overlook noise, temporal consistency, task performance, and computation efficiency.Source: §3.4 · p. 65

3.4.1 Perception-Aware Prediction Evaluation.

  • In joint perception–prediction evaluation, predictions conditioned on detected objects require online matching to ground-truth trajectories, so detection errors or misses can change the trajectory used to assess prediction performance.Protocols differ: DeTra [76] uses an IoU threshold of 0.5, whereas ViP3D [183] uses a distance threshold of 2m with L2 bipartite matching; a unified protocol is still needed.Source: §3.4.1 · pp. 66–67

3.4.2 Planning-Aware Prediction Evaluation.

  • Planning-aware evaluation should measure practical utility, since best-case or marginal trajectory accuracy may not reflect the most likely, safe, or jointly feasible futures planners need.Joint prediction represents interactions and scene-level consistency, while ego-conditioned prediction accounts for how planned actions influence other agents; joint multimodal outputs also raise computation and evaluation challenges.Source: §3.4.2 · pp. 67–68

3.4.3 Closed-loop Evaluation

  • Closed-loop evaluation should measure temporal consistency, inference latency, and task success because open-loop scores may poorly predict deployed system performance.The section recommends latency constraints or reporting and attention to the latency–performance frontier; unstable predictions can disrupt replanning, comfort, safety, and agent interactions, while task-oriented metrics assess ultimate outcomes.Source: §3.4.3 · pp. 69–70

3.5 Perspectives

  • Future benchmarks should bridge isolated prediction and end-to-end evaluation by testing realistic conditions, module interactions, robustness, and practical utility across system configurations.The proposed evaluation agenda includes upstream perception uncertainty, joint perception–prediction assessment, planner-relevant and joint or ego-conditioned prediction, temporal consistency, latency, and task-oriented metrics.Source: §3.5 · p. 71
  • Pruning, lightweight architectures, distillation, and compact designs can reduce runtime demands while balancing accuracy and efficiency for dynamic deployment.The passage describes pruning as reducing computational cost without significant accuracy loss.Source: §3.5 · p. 72
  • Improving efficiency, robustness, and seamless integration is presented as key to advancing motion prediction as a bridge between perception and decision-making.Source: §3.5 · pp. 72–73
  • The outlook is for motion prediction to evolve from a standalone component into a foundation for reliable, adaptable behavior in complex environments.The passage associates this evolution with safer and more efficient deployment across diverse domains.Source: §3.5 · p. 73

4. Generalizable Motion Prediction

Motion prediction must generalize beyond finite, narrow datasets, while data collection for robots and self-driving cars remains costly and incomplete. The survey organizes generalization methods by data access, development stage, and focus, and proposes a lifecycle for adapting across distribution shifts.

Approach taxonomy

  • The survey organizes approaches by target-data access, development stage, and focus, and reviews self-supervision, domain transfer, continual learning, out-of-distribution methods, data synthesis, and foundation models.The covered tasks have distinct settings and use cases but overlapping methodologies.Source: Figure 21 · p. 74, §4 · pp. 76–77, Figure 22 · p. 77

Motivation

  • Finite datasets capture only narrow slices of real-world behavior, while collecting and annotating robot and driving data remains costly and may miss diverse events.Real deployments can expose models to unfamiliar geometries, agents, weather, accidents, and adversarial attacks.Source: §4 · pp. 74–75

Generalization lifecycle

  • The proposed generalization lifecycle cycles through diverse-data collection, universal-model training, safety-aware deployment, and test-time adaptation, then uses detected out-of-distribution data to enrich the next training cycle.The cycle aims to support adaptation across distribution shifts while retaining previously acquired skills where applicable.Source: §4 · p. 75

Self-supervised learning

  • Self-supervised learning adds proxy tasks such as trajectory reconstruction or contrastive learning to learn transferable representations, either through pretraining or alongside supervised prediction.The approach also applies to point-cloud forecasting, video generation, and scene reconstruction.Source: §4 · p. 75

Domain generalization and adaptation

  • Domain generalization tests a source-trained model on a different target domain without target data, whereas domain adaptation uses limited target data to improve target performance without overfitting.Target data may arrive as a small offline batch or online stream, and may lack prediction labels.Source: §4 · p. 75

Continual learning and out-of-distribution methods

  • Continual learning adapts across a sequence of changing domains while seeking to preserve performance on previously seen domains; out-of-distribution methods either detect and reject unusual inputs or predict using additional assumptions or fallback models.Continual learning must avoid forgetting, while out-of-distribution generalization may use representative outliers or route detected inputs to a robust model.Source: §4 · pp. 75–76

Data augmentation and synthesis

  • Augmentation perturbs existing motion examples, while synthesis adds scenarios such as collisions or adversarial agents; both must preserve plausible behavior, map constraints, and interaction dynamics.Motion data’s structural and temporal constraints make realism a central challenge.Source: §4 · p. 76

Foundation models

  • Motion foundation models remain constrained by the size and format discrepancies of public datasets, while industry data raises open questions about distribution measurement, quality, diversity, and balance.Adapting existing foundation models, including language-model reasoning capabilities, is another direction proposed for exploration.Source: §4 · p. 76

4.1 Self-Supervised Learning

The survey organizes self-supervised motion prediction by representation and pretext task, from annotated trajectories to raw sensor streams. It also identifies dataset and evaluation constraints that remain relevant to generalization.

4.1 Self-Supervised Learning; 4.1.1 Self-Supervised Learning for Trajectory Prediction

  • Trajectory SSL improves representations rather than eliminating annotation, and spans joint multitask or pretrain–fine-tune training across five pretext-task families.The families are masking-reconstruction, contrastive learning, redundancy reduction, auxiliary motion-related tasks, and synthetic-trajectory pretraining.Source: §4.1 · p. 78, §4.1.1 · pp. 78–79

4.1.1.1. Masking-Reconstruction

  • Traj-MAE [87] masks inputs before encoding, while Forecast-MAE [101] masks encoded features; both reconstruct inputs and fine-tune the encoder for downstream prediction.Source: §4.1.1.1 · pp. 79–80

4.1.1.2. Contrastive Learning

  • Contrastive methods form trajectory, map, or trajectory-map positive pairs, but mistaken negatives can misalign representations and degrade performance.Works also combine contrastive objectives with other pretext tasks or motion reconstruction.Source: §4.1.1.2 · p. 81

4.1.1.3. Redundancy Reduction; 4.1.1.4. Auxiliary Motion-Related Tasks

  • RedMotion uses redundancy reduction to learn augmentation-invariant road features, while SSL-lanes [50] and Social-SSL use auxiliary tasks for map, agent, or interaction information.Source: §4.1.1.3 · p. 81, §4.1.1.4 · p. 81

4.1.1.5. Prediction on Synthetic Trajectories

  • Synthetic trajectories generated from HD-map geometry can pretrain motion predictors, but simplified motion and interaction patterns create a sim-to-real gap and limit scalability.High-fidelity simulation also requires substantial computation and careful design.Source: §4.1.1.5 · p. 82

4.1.2 Point Cloud Forecasting

  • Point-cloud forecasting predicts future scans from past LiDAR sequences, enabling detection and tracking modules to recover future trajectories without annotated tracks.Forecasting 4D occupancy can also support zero-shot cross-sensor generalization by rendering scans for different LiDAR sensors.Source: §4.1.2 · pp. 82, 84

4.1.3 Video Generation and Scene Reconstruction

  • Driving video generators condition future imagery on controls and can produce consistent multi-view scenes, while 4D reconstruction adds geometry but per-scene optimization can take hours or minutes.Feedforward scene-reconstruction methods replace per-scene optimization with a single pass; early work also explores 4D feedforward reconstruction for driving.Source: §4.1.3 · p. 85

4.1.4 Perspectives

  • Public motion datasets are limited and imbalanced, SSL comparisons lack equivalent-compute controls, and sensor-stream methods remain computationally demanding and limited in capturing discrete objects.Dataset incompatibilities hinder unified SSL pretraining, and the optimal division of a fixed compute budget between pretraining and fine-tuning remains open.Source: §4.1.4 · pp. 85–87

4.2 Domain Generalization and Adaptation

This section surveys approaches for addressing domain shifts in motion prediction, including invariant features, domain-specific modules, causal learning, and few-shot or test-time adaptation. It emphasizes realistic assumptions about target data and standardized cross-dataset evaluation, noting distinctions between limited or streaming labeled data and large-scale unlabeled data, as well as dataset incompatibility.

4.2 Domain Generalization and Adaptation

  • Domain generalization trains on source domains without target data, whereas domain adaptation can use unlabeled or few target examples to address domain shifts.The section groups approaches into invariant-feature design and learning, domain-specific modules, causal learning, and few-shot or test-time adaptation.Source: §4.2 · pp. 87–88

4.2.1 Designing and Learning Invariant Feature

  • Invariant-feature methods reduce sensitivity to domain shifts through designed representations or learned cross-domain alignment.Motion-prediction representations include Frenet coordinates, hierarchical intention and motion features, trajectory transformations, and continuous-time models; learned alignments include direct, adversarial, distillation, and class-conditioned losses. Differences in trajectory length and object count also complicate cross-domain alignment.Source: §4.2.1 · pp. 89–92

4.2.2 Domain-Specific Modules

  • Domain-specific modules capture shifts with small adaptations, often using limited target data while retaining a source-trained model.MoSA [296] adds low-rank encoder connections and outperforms full-model or final-layer fine-tuning; adapting only the encoder corresponding to the dominant shift can avoid overfitting. FLN selects a subnetwork by observed sequence length, while Park et al. use domain-specific noise modules to handle dataset-specific ground-truth errors.Source: §4.2.2 · pp. 92–93

4.2.3 Causal Learning

  • Causal methods seek stable predictive signals by suppressing domain-varying spurious correlations through invariant risk minimization or causal intervention.IRM requires multiple training domains, while many intervention methods can operate with a single domain; Chen et al. [86] subtract counterfactual prediction bias, and GCRL [33] samples latent confounder distributions.Source: §4.2.3 · pp. 94–96
  • Causal-learning methods model non-causal environmental effects with sampled or learned features to make predictions more robust to those effects.SEAD estimates mean environmental effects using learnable vectors, while CaDeT samples non-causal features and minimizes variation in predicted positions across samples.Source: §4.2.3 · p. 96

4.2.4 Few-Shot and Test-Time Adaptation

  • Few-shot and test-time adaptation methods use online optimization, meta-learning, or memory, with the first two updating model parameters and memory methods changing the network inputs.The approaches adapt to new domains with limited data; memory-aided methods require a different encoder architecture.Source: §4.2.4 · pp. 96–97
  • Memory-aided methods retrieve similar past trajectories or predictions as references, with MANTRA retaining high-error cases and OATMem using a FIFO queue.OML-PTP selects memory items by feature-space similarity using multi-hop attention; OATMem also uses domain-specific networks and k-means to maintain a compact shared memory.Source: §4.2.4 · pp. 97–99
  • HATN [569] [571] [568] uses a modified EKF to update weights from errors across past prediction steps, improving short-horizon prediction in previously unseen scenarios.Other online approaches update the final layer with Kalman-filter-like and gradient steps, optimize agent-specific tokens with streaming data, or update graph node encoders and edge weights online.Source: §4.2.4 · pp. 97–98
  • MetaTraj constructs adaptation subtasks from input trajectories and can adapt without target-domain ground-truth labels.Moon and Seo instead learn user-specific embeddings, optimizing them for an individual while updating a shared encoder across users.Source: §4.2.4 · p. 98

4.2.5 Perspectives

  • Target-domain input-only unlabeled data is often an artificial assumption for motion prediction, where input and output trajectories are temporal splits of the same sequence and may be available together.The paper argues that limited target data or streaming labeled data is more meaningful, with the challenge of learning transferable features or adapting efficiently while avoiding overfitting.Source: §4.2.5 · p. 99
  • Cross-dataset evaluation remains difficult because motion-prediction datasets differ in horizons and map formats, and the field lacks an official benchmark.Some methods unify formats and evaluate across datasets, but the paper argues a standardized benchmark is essential for fair comparison and progress assessment.Source: §4.2.5 · p. 99

4.3 Continual Learning

Continual learning adapts models as test distributions change while seeking to preserve performance on previously seen tasks. The section reviews replay and regularization approaches, their motion-prediction applications, and open challenges around task boundaries and online learning.

4.3 Continual Learning

  • Continual learning adapts as test distributions change while retaining old-task performance; motion prediction most often faces domain-incremental shifts in environments or agent behavior.Source: §4.3 · pp. 99–100
  • Replay stores or generates old-task examples, regularization constrains changes to important weights, and parameter isolation has not been used in motion prediction to the authors’ knowledge.Source: §4.3 · p. 100

4.3.1 Replay Methods

  • GEM projects gradients to avoid increasing past-task loss, whereas ER averages training on current- and past-task samples; GSS [18] favors diverse stored examples.Source: §4.3.1 · pp. 100–101, Figure 35 · p. 100
  • Generative replay synthesizes old-task examples instead of storing them; Social-GR eliminates memory by conditioning generation on the new-domain trajectory’s initial waypoint.Source: §4.3.1 · pp. 101–102, Figure 36 · p. 101
  • D-GSM [353] allocates stored examples by scenario divergence, while CLTP-MAN stores 1% of each task’s trajectories and uses a second memory for current-task examples.Source: §4.3.1 · pp. 101–102
  • Kang et al. combine experience replay, meta-learning, and a task mask; Ma et al. generate full trajectories from three stored waypoints, while GRTP conditions on past motion and other vehicles’ relative positions.Source: §4.3.1 · p. 102

4.3.2 Regularization

  • Regularization limits changes to weights important for prior tasks, but EWC [286] can saturate as tasks accumulate; motion-prediction examples combine EWC-like constraints with MMD alignment or replay.Source: §4.3.2 · pp. 102–103

4.3.3 Perspectives

  • Task-aware learning assumes known task boundaries, while task-free learning must detect unannounced changes; online continual learning that preserves past-task performance remains underexplored.Source: §4.3.3 · p. 103

4.4 Out-of-Distribution Detection and Generalization

OOD methods address unfamiliar inputs through detection and robust prediction or fallback, but motion-prediction regression makes shifts harder to define and benchmark. The survey reviews detection and fallback strategies and highlights OOD-aware decision-making and data collection as open directions.

4.4 Out-of-Distribution Detection and Generalization

  • OOD shifts include novel agents, behaviors, environments, and traffic rules; measuring shift severity helps decide between rejecting inputs and applying OOD generalization.OOD generalization aims to maintain reliable predictions for minor discrepancies, while detection can reject inputs that exceed a threshold.Source: §4.4 · pp. 103–104

4.4.1 OOD Detection

  • OOD detection methods use past prediction errors or labeled outliers, predictive uncertainty, or generative likelihood and reconstruction signals.Error-based detection can be delayed until future trajectories become available, while trained inlier-outlier classifiers assume training outliers represent potential test-time outliers.Source: §4.4.1 · pp. 104–106
  • Sun et al. found direct classification of large prediction errors outperformed ensemble disagreement, GAN discrimination, and Bayesian error-likelihood estimation for OOD detection.The authors hypothesize that direct classification captures predictor failure modes more directly than indirect uncertainty estimates.Source: §4.4.1 · p. 106

4.4.2 OOD Generalization

  • OOD generalization combines detection with fallback or fusion, pairing learned predictors with rule-based strategies that trade accuracy and interaction for safer trajectories.Examples include routing between learned and planner-based predictors, and RuleFuser [437], which combines evidence from learned predictions and a rule-based planner.Source: §4.4.2 · pp. 107–108, Figure 38 · p. 107

4.4.3 Perspectives

  • OOD-aware decision-making must balance robust prediction against recognizing when shifts exceed trust, while OOD methods can help collect rare samples for conservative expansion of training data.Fusing detection and generalization remains under-researched; deployment data collection can gradually broaden the data scope while minimizing risk.Source: §4.4.3 · pp. 108–109
  • Regression-based motion prediction lacks fine-grained OOD definitions: treating every sample from a new dataset as OOD can misclassify covered samples and miss underrepresented source-domain regions.Future-waypoint regression shifts emphasis toward input covariate shifts, and many classification-based OOD methods do not apply directly.Source: §4.4.3 · pp. 108–109

4.5 Data Augmentation and Synthesis

The survey covers data augmentation and synthesis as ways to broaden motion-data coverage, while emphasizing motion constraints and unresolved questions about distributional coverage. It reviews augmentation, scenario-generation methods, evaluation, and future directions for data characterization and controllable generation.

4.5.1 Method Recap

  • Motion data augmentation and synthesis must respect dynamics, map layouts, and agent interactions, unlike image manipulations that are comparatively easy to apply.The survey groups methods into augmentation and synthesis and reviews metrics for generated-data quality.Source: §4.5.1 · p. 109, §4.5.2 · p. 110

4.5.2 Application in the Motion Domain

  • Agent and lane dropout, scene rotations or flips, and prediction of contextual agents augment motion data by masking or transforming existing scenes.SceneTransformer treats contextual agents moving more than 6 meters as additional prediction targets during training.Source: §4.5.2 · pp. 110–111
  • Motion scenario synthesis spans rule-based data generation, learned agent policies, and generative models, with approaches ranging from map and traffic construction to multi-agent trajectory generation.TrafficGen [151] autoregressively generates agent initial states and trajectories; CTG conditions diffusion on trajectory attributes, while RealGen uses retrieved scenarios as templates.Source: §4.5.2 · pp. 111–112
  • Adversarial generation creates rare safety-critical scenarios by inserting agents, perturbing existing trajectories, or modeling deployment conditions such as occlusion and perception errors.RobustTraj [70] reports that adversarial training can reduce clean-data performance, while additional augmentations can mitigate this effect.Source: §4.5.2 · pp. 112–113
  • Generated data quality is assessed through realism, diversity, and controllability, but verifying instructions and judging overall quality remain difficult and resource-intensive.Metrics include distributional divergence, rule and map compliance, vehicle dynamics, attribute coverage, and—in LCTGen [528]—human preference judgments.Source: §4.5.2 · pp. 113–114

4.5.3 Perspective

  • The survey calls for better distribution metrics to identify rare or missing scenarios and decompose dataset shifts, and for efficient generation of maps that constrain trajectories and interactions.It also proposes combining controllable closed-loop generation with adversarial agents to produce representative and adversarial scenarios in one simulator.Source: §4.5.3 · pp. 114–115

4.6 Foundation Model - Scale Up the Data and Model

Motion prediction remains data-limited, and integrating datasets is complicated by differing formats, APIs, and prediction settings. The survey reviews early scaling efforts and identifies unresolved questions about pretraining, data utility, and adapting existing foundation models.

4.6 Foundation Model - Scale Up the Data and Model

  • Motion prediction datasets remain small relative to NLP and computer vision, while format, API, horizon, map-resolution, sampling-rate, and annotation differences complicate unified training and evaluation.The survey frames multi-dataset integration as an early step toward scaling motion data, despite costly collection and annotation.Source: §4.6 · p. 115
  • UniTraj [153] reports that models trained on the larger, more diverse Waymo dataset generalize best, and larger models scale more strongly with additional data.SmartPretrain [679] combines datasets through dataset-agnostic sampling and reports consistent performance gains across models, datasets, splits, and main metrics; its contrastive and reconstructive tasks outperform standard motion-prediction tasks at learning informative features.Source: §4.6 · pp. 115–116, Figure 40 · p. 116

4.6.1 Perspectives

  • Open questions include how to learn features that generalize across data sources, measure new data's utility and distribution, and adapt internet-scale language or video models to motion prediction.The survey notes that existing foundation-model adaptation efforts, including motion tokenization, remain nascent.Source: §4.6.1 · pp. 116–117
  • Using 500,000 hours of driving data, a recent study [38] reports predictable power-law performance gains with compute scaling and improving closed-loop metrics; compute-optimal model size grows 1.5× faster than dataset size.The study also reports that sampling and clustering outputs from smaller models can remain competitive up to a crossover point.Source: §4.6.1 · p. 117

4.7 Outlooks

The outlook treats open-world generalization as a continuing lifecycle, while identifying limited and inconsistent data, fragmented evaluation, and immature adaptation methods as persistent challenges. It highlights motion foundation models as a nascent direction whose data, modeling, and deployment questions remain open.

Data limitations

  • Motion prediction remains in a small-data regime, and costly collection plus incompatible dataset formats constrain transferable learning and cross-dataset use.Differences include trajectory horizon, sampling frequency, coordinate system, map modality, and agent taxonomy.Source: §4.7 · p. 117

Benchmarking

  • Motion datasets differ in key formats and definitions, leaving a widely accepted cross-dataset benchmark absent despite preliminary unification efforts.Customized evaluation splits impede fair comparison, reuse of methods, and progress tracking.Source: §4.7 · p. 118

OOD evaluation

  • Classification-style semantic-shift definitions do not directly fit regression prediction, while treating whole unseen datasets as OOD misses within-dataset overlap and rare cases.The survey points to trajectory diversity, map complexity, and interaction density as possible bases for finer OOD definitions.Source: §4.7 · p. 118

Evaluation protocols

  • Generalization evaluations remain fragmented: narrow settings and accuracy-focused metrics often omit uncertainty’s operational role in planning and control.Methods can share mechanisms across tasks, but are often tested at fixed data scales or under one shift type.Source: §4.7 · pp. 118–119

Data understanding

  • A lack of unified motion-data distribution tools complicates data selection, collection, synthesis, and model evaluation across the generalization pipeline.The survey calls for measures that identify coverage gaps, redundancy, and interpretable differences such as interaction complexity or road geometry.Source: §4.7 · pp. 119–120

Adaptation and robust models

  • Transferable representations, online adaptation, and OOD-aware prediction remain open challenges, with current methods often narrow in scope and benchmarks poorly suited to long-term adaptation.The survey calls for clearer OOD definitions, improved uncertainty calibration, and evaluation aligned with downstream needs.Source: §4.7 · p. 120

Foundation models

  • Motion foundation models remain nascent, with open questions spanning dataset unification, motion-specific architectures, modality gaps, and efficient adaptation to downstream tasks.The survey also points to scaling as a possible route to open-world robustness, while noting motion prediction has yet to fully leverage scalable paradigms.Source: §4.7 · pp. 121–122

Open-world generalization

  • Open-world generalization spans data acquisition, learning, runtime uncertainty estimation, and online adaptation, with deployment feeding new OOD data into a continuing cycle.The survey covers approaches including self-supervised learning, domain adaptation and generalization, continual learning, OOD methods, data synthesis, and foundation models.Source: §4.7 · pp. 117, 122

5. Conclusion

The survey revisits motion prediction through deployability and generalizability, noting a gap between benchmark progress and performance in deployed systems or open-world settings. It organizes methods and challenges across autonomy stacks, generalization approaches, and evaluation.

Conclusion

  • Despite benchmark progress, state-of-the-art approaches can fall short in deployed systems and open-world conditions, motivating a rethink of formulations, assumptions, and evaluation.The survey frames deployability and generalizability as under-explored dimensions for real-world adoption.Source: §5 · p. 123
  • The survey presents a taxonomy of representations, modeling paradigms, applications, and evaluation, then examines deployability and methods for generalization.Its generalization coverage includes self-supervision, domain adaptation and generalization, continual learning, OOD methods, data augmentation and synthesis, and foundation models.Source: §5 · p. 123
  • Prediction in autonomy stacks requires defined interfaces, compatible learning signals, and awareness of upstream inputs and downstream requirements; generalization must be deliberately pursued.Unified benchmarking is especially complex because deployability and generalization studies often use differently structured datasets and assumptions.Source: §5 · p. 123
  • The survey aims to guide research beyond idealized benchmarks toward the complexity of real-world autonomy and open-world readiness.This is presented as the intended role of the survey’s roadmap for future motion prediction research.Source: §5 · p. 123
Loading 2505.09074v4…