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

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

cs.ROSep 3, 2026

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

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

Organizations: Georgia Institute of Technology · University of Edinburgh

Abstract

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.

Figures & tables

Fig. 1 : Deadline cliff as more AVs participate in fusion. The baseline fuses every available AV on each update and runs on edge-class hardware. Fusion grows with the number of participating AVs and becomes the largest edge cost beyond four AVs. Stacked bars show mean per-component latency contribution and diamond markers show p95 AoI, crossing the 300 ms bound between 17 AVs (292 ms) and 19 AVs (302 ms).

System	Input	Output
LiveMap [ 1 ] , AdaMap [ 2 ]	detected objects	object map, tracks
Harbor [ 3 ] , C-MASS [ 4 ]	selected perception data	fused detections
Where2Comm [ 5 ]	feature regions	detections, one AV
CMP [ 6 ] , V2XPnP [ 7 ] , Co-MTP [ 8 ]	features, tracks	trajectories, one AV
Conductor	BEV features (vehicles + RSU)	tracks + trajectories, all planners in locale

TABLE I : Task boundaries of the closest systems.

Term	Definition
Locale	RSU-anchored intersection or road segment
Contributor	CAV that uploads features to the edge
Recipient	CAV that consumes the published world model
AoI	World model’s age when the CAV planner reads it
$N$	Cooperating agents in a locale , including the RSU
$K_{\mathrm{cav}}$	Selected CAV contributors for one publish cycle

TABLE II : Key terms and notation.

Panel of Figure 4 — Fig. 2 : Conductor pipeline. Contributors uplink compressed features, confidence maps, and poses; the edge filters to $K_{\mathrm{cav}}$ , runs fusion, detection, tracking, and prediction once per update, and multicasts the world model. Recipients consume it on their 20 Hz planner cycle. The controller (bottom) adapts $K_{\mathrm{cav}}$ and prediction scheduling under the compute SLO $D$ .

Fig. 3 : Envelope compliance (% of updates with composed AoI

{\leq}300

ms) vs. participant count

N

for three policies. The Joint controller is the only policy that complies with the 300 ms safety bound across the full density sweep. Static and Prediction-only both collapse beyond

N{=}16

as fusion overshoots the deadline.

Config	mean	p95	env.	R	R occ
	(ms)	(ms)	(%)
Static-all ( $K{=}N{=}22$ )	233	242	0	0.590	0.447
Random ( $K_{\mathrm{cav}}{=}4$ )	154	155	98	0.587	0.434
Occlusion-aware ( $K_{\mathrm{cav}}{=}4$ )	147	152	100	0.592	0.438
Joint ( $K^{*}_{\mathrm{cav}}{=}4$ , +adapt.)	145	150	100	0.592	0.438

TABLE III : Mechanism ablation for Locale A (

N{=}23

, 50 updates per config). Envelope compliance (%): fraction of cycles with composed AoI inside 300 ms ( Section VI-F ); mean/p95 refer to the edge latency component (not end-to-end latency). Absolute recall is bounded by a model-and-dataset structural ceiling ( Section III ).

Fig. 4 : Total recall (a) and RSU-occluded recall (b) vs.

K_{\mathrm{cav}}

on Locale B (

\bar{N}{=}12

, RSU occlusion

{\sim}45

%). At the high-occlusion locale , four well-selected contributors close 87% of the gap to oracle on RSU-occluded recall. The Occlusion-aware selector tracks the Oracle.

Prediction source	$n$	FDE (m)			ADE (m)
		1 s	3 s	5 s	1 s	3 s	5 s
Fresh MTR	552	3.48	4.19	4.64	2.61	4.04	4.64
Cache (amortized)	495	3.67	4.32	4.66	3.04	4.28	4.83
Linear fallback	53	2.44	8.96	8.92	1.27	4.04	6.43
Stationary const.	5,369	0.61	0.64	0.67	0.61	0.62	0.65

TABLE IV : Prediction quality by source under the Joint controller on Multi-V2X. FDE and ADE are computed against ground truth. For MTR-based rows we use the closest of MTR’s six predicted trajectories, matching CMP’s protocol [ 6 ] ; linear fallback and stationary reuse emit one trajectory. Moving-track rows use

v{\geq}2

m/s; the stationary row uses

v{<}2

m/s.

Fig. 5 : AoI. (a) p95 AoI vs.

N

per policy; reference lines mark the 300 ms safety bound and the 130 ms internal compute SLO. (b) Mean AoI decomposition at

N{=}24

stacked by component. Joint stays inside the bound across

N\in[4,24]

; Static and Prediction-only breach the 300 ms bound at

N{=}24

DL PRR	p95 AoI (ms)	Compliance (%)
1.00	220	100.0
0.95	225	99.9
0.90	231	99.8
0.85	238	99.5

TABLE V : Joint controller’s sensitivity to multicast delivery loss. Drops extend AoI in 50 ms planner-cycle increments because the planner reuses the last received world model while operating at 20 Hz.

Panel of Figure 11 — Fig. 6 : LTAP/OD scenario for closed-loop evaluation. Ego (blue) plans an unprotected left turn against a moving non-AV (red, 13 m/s, distracted). RSU (yellow) on NE pole; cooperating CAVs (green) in unrelated lanes. Visibility axis (a vs. b) toggles three occluders on the diagonal sight-line; load axis ( $N{=}4$ vs. $12$ ) varies queued CAVs (not depicted, no new threat actors).

Fig. 7 : Closed-loop scenario success across the visibility

\times

load matrix. Over-bar numbers: success rate (top) and min-TTC (bottom).

Fig. 8 : Closed-loop success vs. AoI, averaged over 10 runs per (visibility, load, method). Marker size: load (small: low

N{=}4

, large: stress

N{=}12

); shape: method. The right-side shaded region violates the

Z{=}300

ms deadline. Joint is the only method landing inside the envelope with high success under every condition.

Method	LTAP/OD	SCP	Blind overtake
Late fusion	0.40 (1.3)	0.30 (1.0)	0.40 (1.5)
Static cooperative	0.10 (1.0)	0.10 (0.7)	0.10 (1.0)
Joint (ours)	0.90 (2.1)	0.90 (2.0)	0.90 (2.6)
CMP-class per-AV	0.48 (0.5), 21 runs		—

TABLE VI : Occluded, stressed cells across conflict geometries: success rate (min-TTC, s) over 10 seeds. CMP-class row: 21 occluded LTAP/OD+SCP runs.

Explore similar work

Apr 24, 2026cs.RO

Collaborative Trajectory Prediction via Late Fusion

Predicting future trajectories of surrounding traffic agents is critical for safe autonomous navigation and collision avoidance. Despite all advances in the trajectory forecasting realm, the prediction models remains vulnerable to uncertainty caused by occlusions, limited sensing range, and perception errors. Collaborative vehicle-to-vehicle (V2V) approaches help reduce this uncertainty by sharing complementary information. Existing collaborative trajectory prediction methods typically fuse feature maps at the perception stage to construct a holistic scene view. Further this holistic representation is decoded into the future trajectories. Such design incurs substantial communication overhead due to the exchange of high-dimensional feature representations and often assumes idealized bandwidth and synchronization, limiting practical deployment. We address these limitations by shifting collaboration from perception to the prediction module and introducing a late-fusion framework for shared forecasts. The framework is model-agnostic and treats collaborating vehicles as independent asynchronous agents. We evaluate the approach on the OPV2V, V2V4Real, and DeepAccident datasets, comparing individual and collaborative forecasting. Across all datasets, late fusion consistently reduces miss rate and improves trajectory success rate (

\mathrm{TSR}_{0.5}

), defined as the fraction of ground-truth agents with final displacement error below 0.5 m. On the real-world V2V4Real dataset, collaborative prediction improves the success rate by

1.69\%

and

1.22\%

for both intelligent vehicles, respectively, compared with individual forecasting.

Nadya Abdel Madjid, Murad Mebrahtu, Zakhar Yagudin +5

Khalifa University, Computer Science, Abu Dhabi, UAE · Khalifa University, KUCARS-KU Center for Autonomous Robotic Systems, Department of Computer Science, Abu Dhabi, UAE · Khalifa University, Computer and Information Engineering, Abu Dhabi, UAE +2

Multi-Agent Trajectory Prediction V2X Communication Vehicle Trajectory Prediction

Ask PDF

Aug 10, 2026cs.CV

Towards Collaborative Joint Perception and Prediction: Framework, Baseline Evaluation, and Deployment Perspectives

Connected Autonomous Vehicles (CAVs) increasingly exploit Vehicle-to-Everything (V2X) communication to exchange multi-source sensor information, enabling advanced Collaborative Perception (CP) capabilities. Extending beyond these capabilities, this work focuses on Collaborative Joint Perception and Prediction (Co-P&P), a paradigm that unifies CP with motion prediction to mitigate two persistent challenges: the accumulation of perception errors and visual occlusions. We present a conceptual framework for Collaborative Joint Perception and Prediction (Co-P&P) that improves motion prediction of surrounding road users, thereby enhancing situational awareness in complex and dynamic traffic environments. Building upon our preliminary study, this extended version compares the performance of different fusion strategies and establishes baseline performance for a modular design of perception and prediction. Experimental results show that prediction-level fusion leads to a decline in overall system performance compared to detection-level or tracking-level fusion. We further implement a minimal end-to-end Co-P&P prototype that couples collaborative point-cloud sharing via the RENO neural codec with joint detection-forecasting via FutureDet, showing that collaboration improves forecasting accuracy while neural compression preserves this benefit at roughly 34x lower communication bandwidth.

Lei Wan, Hannan Ejaz Keen, Alexey Vinel

XITASO GmbH, Augsburg, 86153, Germany. · Karlsruhe Institute of Technology, Karlsruhe, 76131, Germany.

Cooperative Perception Trajectory Prediction Multimodal Sensor Fusion

Ask PDF

May 13, 2026cs.RO

Causality-Aware End-to-End Autonomous Driving via Ego-Centric Joint Scene Modeling

End-to-end autonomous driving, which bypasses traditional modular pipelines by directly predicting future trajectories from sensor inputs, has recently achieved substantial progress. However, existing methods often overlook the causal inter-dependencies in ego-vehicle planning, ignoring the reciprocal relations between the ego vehicle and surrounding agents. This causal oversight leads to inconsistent and unreliable trajectory predictions, especially in interaction-critical scenarios where ego decisions and neighboring agent behaviors must be reasoned about jointly. To address this limitation, we propose CaAD, a Causality-aware end-to-end Autonomous Driving framework that captures these dependencies within a shared latent scene representation. First, we propose an ego-centric joint-causal modeling module that builds on the marginal prediction branch, and learns causal dependencies between the ego vehicle and interaction-relevant agents. Second, we employ a causality-aware policy alignment stage implemented with joint-mode embeddings to align the stochastic ego policy with planning-oriented closed-loop feedback computed from surrounding traffic and map context. On the Bench2Drive and NAVSIM benchmarks, CaAD demonstrates strong closed-loop planning performance, achieving a Driving Score of 87.53 and Success Rate of 71.81 on Bench2Drive, and a PDMS of 91.1 on NAVSIM. The project page is available at https://moonseokha.github.io/CaAD/.

Seokha Moon, Minseung Lee, Joon Seo +2

Korea University · Kakao Mobility

Causal Representation Learning End-to-End Autonomous Driving Multi-Agent Trajectory Prediction Autonomous Driving Planning

Ask PDF