PaperlayerView exact revision on arXiv

Source-linked AI summary

Scalable Edge-assisted Fusion and Path Prediction for Connected Autonomous Vehicles

Tyler Landle, Jackson Isenberg, Abhijit Chatterjee, Alexandros Daglis, Umakishore Ramachandran

arXiv:2609.04364v1cs.ROcs.SYeess.SY

TL;DR

Occlusions limit onboard sensing, while edge-generated world models and motion predictions are useful only if they reach planners within the AoI budget. Conductor builds a shared RSU-anchored model and jointly schedules which CAV inputs to fuse and how much prediction to run. It meets the AoI bound across scenarios with up to 31 CAVs and closes 87% of the random-to-Oracle gap in a highly occluded locale.

  • Problem

    Onboard sensors have limited range and line of sight, and shared edge-generated world models and predictions must reach planners while sufficiently fresh.

    Source: §I · p. 1

  • Method

    Conductor builds a shared RSU-anchored world model, selecting CAVs that add evidence beyond the RSU and adapting fusion and prediction work jointly.

    Source: §I · p. 2, §III · p. 3, C. RSU-Anchored Intermediate Fusion · p. 5

  • Results

    Across traffic scenarios with up to 31 CAVs, Conductor meets the AoI bound; selecting four CAVs closes 87% of the random-to-Oracle gap in a highly occluded locale.

    Source: §I · p. 2, §VIII · p. 14

  • Takeaways & Limitations

    Conductor preserves cooperative sensing benefits under both occlusion and dense load in the evaluated scenarios.

    Source: §VIII · p. 14

    Most density results use offline profiler replay, and behavior beyond the measured Multi-V2X density range is not measured.

    Source: §VII · p. 13

Abstract

from arXiv · show

The planning algorithms inside an Autonomous Vehicle (AV) rely on information from on-board sensors whose line of sight is limited by emerging traffic conditions and occlusions. Edge-assisted creation of a unified world model fusing information from AVs and Road Side Units (RSUs) in a geographical locale, and the prediction of AVs' future trajectories, can enhance the planning algorithms inside AVs to improve quality metrics, such as better traffic flow and collision prevention. AVs participating in such enhancements are called Connected Autonomous Vehicles (CAVs). However, such information generated by the edge (world model and motion predictions) must reach the planners within a tight Age of Information (AoI) time budget to be useful. The state of the art fuses per-CAV information: each AV fuses inputs from other actors locally, which limits both scalability with actor count and quality of results. We present Conductor, an edge-based solution for creating a unified world model from the perspective of a fixed anchor (e.g., an RSU) in a locale and predicting future trajectories of AVs in that locale. Our solution adheres to the AoI time budget by dynamically limiting the number of AVs that would lead to the best quality of results. Specifically, we introduce an occlusion-aware selector that favors information contribution by AVs that detect objects in the locale not covered by RSUs. We pair this selector with a runtime controller that adapts both the number of AV inputs to fuse and the amount of trajectory predictions in each cycle to stay within the AoI time budget. Evaluation on CAV simulation infrastructure shows our joint selector-controller meets the AoI safety bound across traffic scenarios with up to 31 CAVs, with fusion fidelity close to an Oracle and much better than a random selector under the same AoI constraint.

I. INTRODUCTION

Conductor addresses the freshness challenge of shared perception by building one RSU-anchored world model and jointly scheduling AV selection and trajectory prediction. Its evaluation reports AoI compliance across traffic scenarios with up to 31 CAVs and improved occlusion recovery over random selection.

I. INTRODUCTION

  • An edge service fuses compact intermediate features once into an RSU-anchored world model and sends the same update to all CAVs in the locale.This avoids duplicate per-AV fusion while making freshness a system-level dependency for every recipient.Source: §I · p. 1
  • The occlusion-aware selector favors CAVs whose detections add evidence beyond the RSU’s view, while runtime control adapts fusion and prediction work to the AoI budget.The selector uses detector confidence maps already produced by the fusion model.Source: §I · p. 2
  • Conductor treats AV selection and trajectory prediction as one scheduling problem because both consume the shared AoI budget.More AV inputs can improve perception but leave less time for prediction; additional fresh predictions help only if the world model reaches the planner on time.Source: §I · p. 2
  • Across traffic scenarios with up to 31 CAVs, Conductor consistently meets the AoI bound; selecting four CAVs closes 87% of the random-to-Oracle gap in a highly occluded locale.In closed-loop evaluations, late fusion fails in occluded scenarios, while using all available CAVs for feature-level fusion fails in dense scenarios.Source: §I · p. 2

II. RELATED WORK

  • Conductor combines cooperative perception, tracking, prediction, and deadline-aware edge scheduling, unlike prior systems that address components or compose them without a high-actor-count end-to-end deadline.The paper describes Conductor as an edge service for cooperative motion prediction, not a new cooperative perception model.Source: §II · p. 2

A. Cooperative Perception

Conductor places cooperative perception, tracking, and motion prediction in one edge-scheduled service with an explicit AoI objective. It builds one locale-wide model and shares the update with CAV planners.

A. Cooperative Perception

  • Conductor uses confidence maps to select CAVs that add value under a compute deadline, rather than optimizing communication efficiency or detection accuracy alone.Feature-level fusion preserves spatial evidence while avoiding raw point-cloud uploads; late fusion cannot recover objects no individual sensor detects confidently.Source: A. Cooperative Perception · p. 3

B. Cooperative Motion Prediction

  • Conductor centralizes the perception-tracking-prediction stack at the edge under an AoI deadline, extending prior motion-prediction pipelines evaluated at smaller CAV counts.CMP and CoPnP limit evaluation scale to 2–7 CAVs, while Conductor evaluates dense locales with up to 31 CAVs.Source: B. Cooperative Motion Prediction · p. 3

C. Edge Compute Scheduling for Autonomy

  • Unlike request-by-request or single-vehicle adaptive compute, Conductor schedules interdependent fusion, tracking, and prediction stages as one edge service under an AoI constraint.Fusion cost depends on participating CAVs, tracking cost on detection count, and prediction cost on track count.Source: C. Edge Compute Scheduling for Autonomy · p. 3

III. SYSTEM ARCHITECTURE AND DESIGN

  • The service output is a timestamped, RSU-anchored locale scene with persistent 3D tracks, motion state, freshness metadata, and predicted trajectories, shared with every CAV.The edge selects CAV inputs, fuses features, updates tracks, predicts trajectories, and sends one multicast update per cycle.Source: §III · pp. 3–4

A. Design Principles

Conductor’s architecture centers on one persistent, RSU-anchored world model that reaches all locale CAVs within an AoI bound. Intermediate fusion and a staged edge pipeline support that shared service.

A. Design Principles

  • Conductor fuses selected CAV features with the RSU input once, maintaining shared track identity in the RSU frame and designing edge computation to remain bounded as the locale grows.The design uses a 300 ms AoI upper bound, derived from a 4 m braking margin at 14 m/s and validated in closed-loop conflict experiments.Source: A. Design Principles · p. 4

B. Service Boundary and Multicast Scaling

  • Because every CAV consumes identical RSU-anchored state, the edge multicasts one world-model update instead of scheduling separate downlink transmissions per recipient.The input comprises feature, confidence-map, and pose tuples from CAVs and the RSU; the output contains current tracks and short-horizon predictions.Source: B. Service Boundary and Multicast… · p. 4

C. RSU-Anchored Intermediate Fusion

  • Intermediate fusion compresses post-encoder features to roughly 17 KB per CAV while preserving spatial structure, and the RSU frame provides persistent coordinates for tracks and predictions.The RSU’s traffic-light-mounted LiDAR at 4–6 m elevation covers most of the intersection center, reducing the need for CAVs that duplicate its view.Source: C. RSU-Anchored Intermediate Fusion · pp. 4–5

D. Edge Pipeline

  • At each update, the edge filters CAV inputs, fuses selected features, detects objects, updates tracks, predicts trajectories, and multicasts the world-model update for planner use.A bounded jitter buffer orders inputs by source timestamp; the batched Kalman tracker update provides a 4–5× speedup at the evaluated deployment scale.Source: D. Edge Pipeline · p. 5

E. Runtime Control Knobs

Conductor jointly schedules CAV fusion and trajectory prediction under an AoI bound. Its controller adapts the contributors and tracks receiving fresh prediction to meet a compute budget.

E. Runtime Control Knobs

  • The edge targets a 130 ms compute SLO, the residual budget after network delay and planner wait are subtracted from the 300 ms AoI bound.The target applies at high N.Source: E. Runtime Control Knobs · p. 5
  • Fusing more CAVs improves scene quality but leaves less prediction budget, while more predictions help only if information reaches the planner on time.The controller maximizes estimated prediction utility while keeping estimated edge cost within the compute SLO.Source: E. Runtime Control Knobs · p. 5

IV. ADAPTIVE CONTROLLER

  • The controller jointly chooses Kcav, the number of CAVs fused, and x, the tracked objects receiving fresh learned prediction.Other tracks reuse cached predictions or fall back to a lightweight linear predictor.Source: §IV · p. 5, E. Runtime Control Knobs · p. 5
  • The per-update cost model sums selection, fusion, detection, tracking, prediction, and fixed controller-overhead costs.Stage costs are fit on profiled hardware.Source: §IV · p. 6
  • Before scheduling, the controller estimates current track count from selected contributors using a rolling tracks-per-contributor ratio plus slack.The ratio uses 20 updates; slack is the larger of 10% of the estimate or 2 tracks.Source: §IV · p. 6

A. Mechanism 1: Occlusion-Aware Contributor Selec-

Conductor always includes the RSU and selects CAV contributors for the additional detection evidence they provide beyond the RSU’s view. Its confidence-map scoring reuses information already produced for fusion.

A. Mechanism 1: Occlusion-Aware Contributor Selection

  • The RSU is always included; when all candidate CAVs exceed the feasible count, the controller chooses contributors whose detector evidence complements the RSU’s view.The selector can use each contributor’s spatial confidence map, which is already uploaded with the feature tensor and adds no inference.Source: A. Mechanism 1: Occlusion-Aware… · p. 6
  • A CAV’s selection score is high when its confidence map indicates an object where the RSU’s map does not, with candidates beyond the configurable range dropped.The default distance gate is 50 m.Source: A. Mechanism 1: Occlusion-Aware… · p. 6
  • The controller searches Kcav values of 0, 2, 4, 8, or 12 and selects the value maximizing estimated utility under the cost model.Utility estimates how many tracks can receive fresh prediction within the remaining prediction budget.Source: A. Mechanism 1: Occlusion-Aware… · p. 6

B. Mechanism 2: Prediction Cost Amortization

Conductor amortizes prediction by refreshing cached trajectories selectively, then allocates remaining prediction budget among divergent objects by a risk score. The evaluation uses offline profiling and a closed-loop simulation platform.

B. Mechanism 2: Prediction Cost Amortization

  • A max cache age forces refresh, while divergence checks also account for position, heading, and velocity-normalized drift.The velocity-normalized threshold allows up to 100 ms of unmodeled drift; the heading threshold is about 10° and the displacement floor is 0.5 m.Source: B. Mechanism 2: Prediction Cost… · p. 6

C. Mechanism 3: Risk-Budgeted Scheduling

  • The risk score prioritizes fresh prediction using speed, distance to the nearest CAV, occlusion weighting, and closest-approach clearance over a 5 s horizon.It is a scheduling heuristic, not a verified collision-risk estimate; no planner path is uploaded.Source: C. Mechanism 3: Risk-Budgeted Scheduling · p. 7
  • The scheduler ranks divergent objects by risk and greedily assigns fresh MTR prediction until the residual budget is exhausted; others use a lightweight linear predictor.That predictor extrapolates future motion from recent tracked state.Source: C. Mechanism 3: Risk-Budgeted Scheduling · pp. 6–7

A. Experimental Platform

  • The offline profiler replays Multi-V2X point clouds and combines measured edge compute with simulated network delay to produce AoI distributions.It runs WorldFusion, AB3DMOT, and MTR end to end on an A10 GPU.Source: A. Experimental Platform · p. 7
  • The closed-loop platform runs the full eCAV stack with CARLA simulation and live ns-3-mediated uplink and downlink.Both platforms share the edge stack, A10 hardware, and ns-3 5G NR network model.Source: A. Experimental Platform · p. 7
  • The planner runs at 20 Hz, while adaptive WorldFusion runs at 10 Hz; the late-fusion baseline publishes every 50 ms planner cycle.The edge service uses an NVIDIA A10 with 24 GB VRAM.Source: A. Experimental Platform · p. 7

B. Dataset

The evaluation uses Multi-V2X locales with varied agent connectivity and RSU occlusion, then compares selector and controller configurations. The participant-count sweep includes RSU-only fusion through all-CAV fusion.

B. Dataset

  • Multi-V2X covers 6 towns and 56 RSU-anchored locales, with connected-agent counts averaging 11.3 and reaching 33.The reported cases contrast dense, low-occlusion Locale A with moderate-traffic, high-occlusion Locale B.Source: B. Dataset · p. 7
  • Locale A averages 23 agents with 9% RSU occlusion, while Locale B averages 12 agents with 45% occlusion.The paper reports cross-density behavior as an aggregate over the full Multi-V2X dense-locale set.Source: B. Dataset · p. 7

C. Participant Scaling

  • Participant scaling sweeps Kcav across 0, 2, 4, 8, and N−1, with the RSU always included and 50 steady-state frames profiled per setting.Kcav=0 uses only the RSU; Kcav=N−1 fuses all CAVs.Source: C. Participant Scaling · p. 7

D. Baselines and Selectors

  • Seven configurations isolate contributor handling and adaptive prediction, including static-all, random, Oracle, occlusion-aware, filter-only, prediction-only, and joint control.Static-all exposes the deadline cliff; Oracle is an offline upper bound and is not deployable.Source: D. Baselines and Selectors · p. 7
  • The joint controller combines all three mechanisms, while filter-only and prediction-only configurations use subsets of those mechanisms.Random selection is uniform; occlusion-aware selection uses fixed Kcav.Source: D. Baselines and Selectors · p. 7

E. Metrics

The section defines the evaluation metrics and AoI thresholds, then explains how contributor-driven fusion cost causes a deadline cliff. Across the full density sweep, the Joint controller alone maintains AoI compliance.

E. Metrics

  • The evaluation tracks per-stage and total edge latency, 300 ms AoI envelope compliance, 130 ms SLO overruns, detection recall and precision, RSU-occluded recall, MOTA, and ID switches.RSU-occluded recall isolates contributor value beyond the RSU baseline; detections use a 1.5 m center-distance matching threshold.Source: E. Metrics · p. 8

F. AoI Model and Thresholds

  • The 300 ms safety bound is the largest world-model age for which RSS-style emergency braking completes before the focal CAV enters the conflict zone; the 130 ms compute SLO is a controller target, not a safety claim.At N=24, network and consume terms contribute about 130 ms at p95; the model also uses multicast downlink, calibrated uplink delays, and serialization, deserialization, edge-compute, and planner-lag terms.Source: F. AoI Model and Thresholds · p. 8

VI. EVALUATION

  • The evaluation characterizes the deadline cliff, isolates selector and prediction-scheduler effects on quality, and tests whether AoI and perception effects translate into closed-loop driving outcomes.The evaluation roadmap spans density behavior, detection, tracking and prediction quality, and closed-loop outcomes under occlusion and edge load.Source: §VI · p. 8

A. Deadline Cliff Characterization

  • p95 AoI crosses 300 ms between Kcav=17 and 19 as fusion becomes the dominant edge cost; only Joint maintains compliance across the full density sweep.Fusion grows from 10 ms at Kcav=0 to 114 ms at Kcav=22 on Locale A; detection, tracking, and prediction are largely insensitive to contributor count.Source: A. Deadline Cliff Characterization · pp. 8–9, Figure 3 · p. 9

B. Cross-Scale Controller Behavior

The Joint controller preserves AoI compliance across participant densities, while the occlusion-aware selector improves recall where the RSU has blind spots. Its recall benefit depends on locale geometry.

B. Cross-Scale Controller Behavior

  • Joint maintains 100% AoI compliance across the density sweep; at N=24, Static reaches 87% and Prediction-only 84%.Filter-only matches Joint on compliance in this experiment because both select the same budget-feasible contributor count; Joint also adapts prediction work.Source: B. Cross-Scale Controller Behavior · p. 9

C. Contributor Selection Impact

  • At Locale A, where RSU occlusion is 9%, all Kcav=4 selection policies tie on RSU-occluded recall within sampling noise at 0.43–0.44; reducing fusion to K=5 restores feasibility regardless of which four CAVs are chosen.Here the selector contributes compute-cost reduction rather than a recall gain.Source: C. Contributor Selection Impact · p. 9
  • At Kcav=4, Occlusion-aware reaches 0.354 RSU-occluded recall and closes 87% of the Random-to-Oracle gap on high-occlusion Locale B.Random reaches 0.283 and Oracle 0.365; fusing all N−1 CAVs reaches 0.349. The selector slightly exceeds Static-all by excluding noisy CAVs.Source: C. Contributor Selection Impact · p. 9, Figure 4 · p. 9
  • Joint and Occlusion-aware have identical recall because prediction does not change detection recall; Joint’s advantage is AoI compliance, while selection improves recall when the RSU has blind spots.The selector reduces compute everywhere, but its recall contribution depends on locale geometry.Source: C. Contributor Selection Impact · p. 9

D. Mechanism Ablation

Conductor combines contributor selection with adaptive prediction to keep AoI within its deadline while preserving tracking quality. Across the evaluated density range, the Joint controller meets the 300 ms bound; the ablation and prediction results explain how it manages compute and forecast quality.

D. Mechanism Ablation

  • At N=23, reducing Kcav from 22 to 4 cuts edge p95 latency to ≤155 ms and raises envelope compliance from 0% to 98–100%, with essentially unchanged total recall.The ablation attributes the compute reduction to selecting fewer CAVs, which removes the dominant fusion cost.Source: D. Mechanism Ablation · p. 10
  • At N=23, adding adaptive prediction to Filter-only saves 2 ms; fusion still costs ∼43 ms, while pairing prediction adaptation with selection holds compliance at 100% across density.Prediction-only reaches 84% compliance at N=24.Source: D. Mechanism Ablation · p. 10

E. Tracking and Prediction Quality

  • Cache reuse tracks fresh MTR closely, linear fallback is used on <1% of predictions, and stationary-track location prediction remains below 1 m error across all horizons.Linear fallback approaches 2× MTR’s FDE at 5 s, while cache reuse is within 0.02 m at 5 s.Source: E. Tracking and Prediction Quality · p. 10

F. AoI

  • AoI measurement combines edge profiling, ns-3 uplink/downlink samples, and on-device serialization and consume-lag calibration, retaining compute and network variation.The Joint policy’s latency plateaus beyond roughly N=16 as Kcav reaches the largest feasible value at the budget boundary.Source: F. AoI · pp. 10–11
  • The Joint controller keeps p95 AoI within the 300 ms bound across N=4–24, with 100% envelope compliance and p95 AoI of 234 ms at N=24.Static and Prediction-only breach the bound at N=24.Source: Figure 5 · p. 10, F. AoI · p. 11
  • At N=24, Static edge compute alone saturates the budget; Joint meets the bound by reducing absolute fusion cost, with the other latency components contributing smaller shares.Under Joint, those components plus reduced edge compute remain inside the 300 ms envelope.Source: F. AoI · p. 11
  • With multicast PRR 0.85, p95 AoI increases by 18 ms and envelope compliance remains 99.5%, while per-recipient unicast reaches 301 ms p95 at N=24.Each multicast miss extends AoI by 50 ms because the planner reuses the last update for a cycle.Source: F. AoI · p. 11, Table V · p. 11

G. Multicast Sensitivity

The evaluation examines multicast sensitivity, selection across locales, and closed-loop driving under visibility and load variation. It reports continued AoI compliance under tested downlink assumptions and shows why both cooperative perception and deadline-aware control matter.

G. Multicast Sensitivity

  • At N=24, p95 AoI is 244 ms with a 20 ms downlink assumption and 264 ms with 40 ms, with 100% sample-level compliance across 188 samples.Full-trace compliance is 99.99%; the baseline multicast downlink is modeled analytically because the public ns-3 NR module lacks NR-MBS support.Source: G. Multicast Sensitivity · p. 11

H. Selection Beyond One Locale

  • Across 16 locales meeting the participation threshold, the selector closes a median 32% of the random-to-Oracle occluded-recall gap at Kcav=4 and 22% at Kcav=2.Locales where the RSU already covers the scene have no recall gap to close, and two low-occluded-object locales underperform random selection at Kcav=4.Source: H. Selection Beyond One Locale · p. 11

I. Closed-Loop Evaluation

  • Closed-loop tests distinguish fast late fusion that misses occluded threats from static fusion that becomes stale under load; Joint preserves the perception benefit while keeping AoI inside the bound.The evaluation varies visibility and load across three methods and defines operational success using collision avoidance, no false braking, and progress.Source: §I · p. 11, 1) Operational success and the… · p. 11
  • The LTAP/OD evaluation varies visibility and queued-CAV load while holding weather, time of day, target speed, and RSU pose constant.The scenario uses an opposing non-AV crossing vehicle and tests N=4 versus N=12 cooperating CAVs.Source: Figure 6 · p. 12

1) Operational success and the physical safety envelope

  • The physical safety envelope is set at Z=300 ms, near the measured transition from ∼1.0 operational success at AoI ≤280 ms to 0 above 380 ms.The envelope is measured by sweeping delivery delay on an Oracle baseline using ground-truth detections.Source: 1) Operational success and the… · pp. 11–12

2) Scenario family

The scenario family tests whether the cooperative perception advantage survives increasing load and occlusion. Joint retains high success across conditions, while static fusion loses success under stress and late fusion struggles when threats are occluded.

2) Scenario family

  • The scenario variants change crossing-vehicle spawn timing and queued-CAV initial poses to produce AoI variation while preserving the scenario geometry.Visibility toggles parked-vehicle occlusion, and load varies the number of cooperating CAVs queued at the signal.Source: 2) Scenario family · p. 12
  • With occlusion but low load, late fusion has 0.40 success while Static and Joint each reach 1.00; without occlusion or load stress, all three methods reach 1.00.Intermediate fusion combines partial views to recover threats missed by late fusion.Source: 2) Scenario family · p. 12
  • In the occluded-stressed cell, Joint reaches 0.90 success, versus 0.40 for late fusion and 0.10 for Static; it is the only method with high success in that condition.Across conditions, Joint stays at ∼120–150 ms and achieves ≥90% success.Source: 2) Scenario family · pp. 12–13, Figure 7 · p. 12, Figure 8 · p. 13
  • Joint maintains min-TTC of 2.1–2.4 s in every cell, while Static falls to ∼1.0 s in both stressed cells and late fusion reaches 1.3–1.4 s in occluded cells.The figure’s over-bar values encode success rate and min-TTC.Source: 2) Scenario family · p. 12, Figure 7 · p. 12
  • Under stressed load, Static crosses the deadline at 345 and 351 ms and its success falls to 0.10–0.20, regardless of visibility.Late fusion stays within the deadline, but its occluded-cell failures are attributed to unresolved occlusion rather than latency.Source: 2) Scenario family · p. 12

J. Additional Conflict Geometries and a Per-Ego Base-

Across additional conflict geometries, Joint preserved closed-loop success under occlusion and dense load, while the per-AV arm collided frequently. The discussion bounds these results by hardware, speed, evaluation, network, and perception constraints.

J. Additional Conflict Geometries and a Per-Ego Base-

  • Across 21 occluded LTAP/OD and SCP runs, the per-AV arm collided in 11; Conductor succeeded in 9/10 runs in each geometry.The per-AV vehicle-centered canvas sees perpendicular traffic only about 2.7 seconds before conflict at 14 m/s, leaving little time for detection, track confirmation, and prediction history.Source: J. Additional Conflict Geometries and… · p. 13, Table VI · p. 13
  • In occluded stressed cells, Joint succeeded in 9/10 runs in both SCP and blind-overtake, versus late fusion at 3/10 and 4/10 and Static at 1/10 in each.In unobstructed stressed cells, Static reached 1/10 and 2/10, while Joint reached 9/10 and 10/10; the deadline failure was independent of visibility.Source: Table VI · p. 13, J. Additional Conflict Geometries and… · p. 13

K. Failure Modes

  • Service failures appear to planners as missing threats, stale state, or wrong trajectories; Conductor reduces admission and staleness failures but does not guarantee collision freedom.Detection or admission failure can produce missing threats; compute or communication delay produces stale state, while prediction, pose, or timestamp errors can produce wrong trajectories.Source: K. Failure Modes · p. 13

VII. DISCUSSION AND LIMITATIONS

  • The measured deadline cliff is specific to the NVIDIA A10 24 GB evaluation hardware, and the 300 ms AoI bound assumes an urban approach speed of 14 m/s.Lower-end accelerators would select fewer CAVs or lower update rates; at 28 m/s, one lane width of braking reserve corresponds to about 143 ms, requiring fewer CAVs and less fresh prediction.Source: §VII · p. 13
  • Most density-scaling results come from offline profiler replay and do not close the full detection-to-planner loop; behavior beyond the Multi-V2X density range is unmeasured.Closed-loop experiments validate the main safety effect only on representative scenarios.Source: §VII · p. 13
  • The controller uses a fixed 130 ms compute target and does not yet adapt that margin to recent network delay.A deployment could reserve more compute margin when network delay tails grow, or reduce the fusion set when late packets risk making an update stale.Source: §VII · pp. 13–14
  • WorldFusion reaches AP@0.3=0.718, AP@0.5=0.674, and AP@0.7=0.531 on 3,504 Multi-V2X car-only frames; detection misses limit downstream tracking and prediction.On V2X-Sim 2.0, the same architecture reaches AP@0.3=0.753, AP@0.5=0.749, and AP@0.7=0.685, comparable to CoBEVT.Source: §VII · p. 14

VIII. CONCLUSION

  • Conductor kept AoI within 300 ms across the density sweep, closed 87% of the random-to-Oracle gap in the high-occlusion locale, and preserved cooperative-sensing benefits under occlusion and dense load.The conclusion identifies freshness as the limiting systems problem when fusion, tracking, and prediction run together at the edge.Source: §VIII · p. 14

Related papers

Loading 2609.04364v1…