HyDRA: Hybrid Domain-Aware Robust Architecture for Heterogeneous Collaborative Perception

Minwoo Song; Minhee Kang; Heejin Ahn

cs.CVMar 25, 2026

HyDRA: Hybrid Domain-Aware Robust Architecture for Heterogeneous Collaborative Perception

Authors: Minwoo Song, Minhee Kang, Heejin Ahn

Organizations: School of Electrical Engineering, Korea Advanced Institute of Science and Technology, Daejeon, Republic of Korea

Abstract

In collaborative perception, an agent's performance can be degraded by heterogeneity arising from differences in model architecture or training data distributions. To address this challenge, we propose HyDRA (Hybrid Domain-Aware Robust Architecture), a unified pipeline that integrates intermediate and late fusion within a domain-aware framework. We introduce a lightweight domain classifier that dynamically identifies heterogeneous agents and assigns them to the late-fusion branch. Furthermore, we propose anchor-guided pose graph optimization to mitigate localization errors inherent in late fusion, leveraging reliable detections from intermediate fusion as fixed spatial anchors. Extensive experiments demonstrate that, despite requiring no additional training, HyDRA achieves performance comparable to state-of-the-art heterogeneity-aware CP methods. Importantly, this performance is maintained as the number of collaborating agents increases, enabling zero-cost scaling without retraining.

Figures & tables

Fig. 1: Heterogeneity in dynamic collaboration. (b) Architecture Heterogeneity occurs when agents employ diverse model architectures, making direct feature fusion structurally infeasible. (c) Latent Domain Heterogeneity is from an independent training setting even when agents adopt identical architectures, resulting in a more subtle threat. Our proposed framework successfully identifies not only explicit structural mismatches but also these hidden latent domain shifts to prevent feature contamination.

Fig. 2: Overview of HyDRA . Domain Classifier computes a domain similarity score $\mathcal{S}_{\text{domain}}$ by comparing the ego-decoded prediction $\mathcal{B}_{pred}$ with the received auxiliary detection result $\mathcal{B}_{A}$ . In Hybrid Fusion , homogeneous agents ( $\mathcal{N}_{int}$ ) participate in intermediate fusion to generate Stage 1 detections ( $\mathcal{B}_{stage1}$ ), while heterogeneous agents ( $\mathcal{N}_{late}$ ) participate in late fusion using their transmitted detection results. At the late-fusion stage, Anchor-Guided Pose Graph Optimization mitigates pose noise by treating the reliable Stage 1 results as fixed spatial anchors to correct the poses of heterogeneous agents.

Fig. 3: Concept of AG-PGO . Instead of global optimization, we leverage the reliable detections from homogeneous agents ( $\mathcal{N}_{int}$ ) as fixed spatial anchors ( ). The module corrects the pose of heterogeneous agents ( $\mathcal{N}_{late}$ ) by minimizing the residual errors between their predictions and the anchors, ensuring global consistency.

(a) Ideal Setting (No Noise)
Method	[email protected]				[email protected]				[email protected]
Method	vehicle	pedestrian	truck	total	vehicle	pedestrian	truck	total	vehicle	pedestrian	truck	total
No Fusion	0.6471	0.3285	0.5161	0.4972	0.6006	0.1248	0.4300	0.3851	0.3208	0.0016	0.2306	0.1843
Late Fusion	0.7297	0.4327	0.4898	0.5507	0.7096	0.1742	0.4167	0.4335	0.4241	0.0073	0.2654	0.2323
E2E Training	0.8641	0.4087	0.5542	0.6090	0.8438	0.1444	0.5038	0.4974	0.5693	0.0031	0.2547	0.2757
MPDA [ 8 ]	0.8288	0.3790	0.5645	0.5908	0.8061	0.1463	0.5426	0.4983	0.5241	0.0045	0.3850	0.3045

TABLE I: Performance comparison under architecture heterogeneity :

E_{P}+H_{P}+X_{S}+X_{V}

. Best and second-best results are highlighted in bold and underline , respectively.

Fig. 4: Performance comparison with baseline methods under varying pose noise and scalability analysis

Method	[email protected]				[email protected]				[email protected]
Method	vehicle	pedestrian	truck	total	vehicle	pedestrian	truck	total	vehicle	pedestrian	truck	total
No Fusion	0.6471	0.3285	0.5161	0.4972	0.6006	0.1248	0.4300	0.3851	0.3208	0.0016	0.2306	0.1843
Late Fusion	0.8399	0.4020	0.4168	0.5529	0.8082	0.1680	0.3706	0.4489	0.5109	0.0074	0.2624	0.2603
Intermediate Fusion	0.7515	0.3554	0.4863	0.5311	0.7195	0.1341	0.4695	0.4410	0.4639	0.0033	0.2434	0.2369
HyDRA (Ours)	0.8774	0.4390	0.5883	0.6349	0.8665	0.1753	0.5003	0.5141	0.5629	0.0056	0.4255	0.3313

TABLE II: Performance comparison under latent domain heterogeneity :

E_{P}+H_{P}+X_{P}+X_{P}

Agent Type	Noise Setup	Domain Score ( $S_{\text{domain}}$ )
Agent Type	Noise Setup	Mean	Max	Min
$H_{P}$ (Hom.)	w/o Noise	0.5136	0.7046	0.2458
$H_{P}$ (Hom.)	w/ Noise	0.5135	0.7013	0.2361
$X_{P}$ (Het.)	w/o Noise	0.0072	0.1444	0.0000
$X_{P}$ (Het.)	w/ Noise	0.0070	0.1421	0.0000
$X_{S}$ (Het.)	w/o Noise	0.0001	0.0134	0.0000

TABLE III: Analysis of Domain Classifier Scores

Fig. 5: Analysis of domain classifier scores comparing homogeneous vs. heterogeneous agents under ideal and noisy (

\sigma=0.4

) conditions.

Domain Classifier	AG-PGO	[email protected]	[email protected]	[email protected]
-	-	0.2559	0.1173	0.0106
-	✓	0.2827	0.1627	0.0245
✓	-	0.4886	0.3829	0.1568
✓	✓	0.5093	0.3904	0.2164

TABLE IV: Ablation Study under Pose Noise Error (

\sigma=0.4

)

Fig. 6: Qualitative comparison of 3D detection results with baseline methods. The red and green boxes represent the prediction and ground truth.

Explore similar work

HeteroPROMPT: A Real-time and Privacy-Preserving Heterogeneous Collaborative Perception Framework
Jul 28, 2026Armin Maleki, Hayder RadhaAutonomous Driving Perception Cooperative Perception
Electrical and Computer Engineering, Michigan State University
GT-Space: Enhancing Heterogeneous Collaborative Perception with Ground Truth Feature Space
Mar 13, 2026Wentao Wang, Haoran Xu, Guang TanHeterogeneous Collaborative Perception Autonomous Driving Perception
Sun Yat-sen University, Shenzhen Campus, Shenzhen, China · Peng Cheng Laboratory (PCL), Shenzhen, China
Faster-HEAL: An Efficient and Privacy-Preserving Collaborative Perception Framework for Heterogeneous Autonomous Vehicles
Mar 7, 2026Armin Maleki, Hayder RadhaAutonomous Driving Perception Heterogeneous Collaborative Perception
Electrical and Computer Engineering, Michigan State University, East Lansing, Michigan

Jul 28, 2026J/K move · Enter open · S save

HeteroPROMPT: A Real-time and Privacy-Preserving Heterogeneous Collaborative Perception Framework

Armin Maleki, Hayder Radha

Electrical and Computer Engineering, Michigan State University

Autonomous Driving Perception Cooperative Perception Heterogeneous Collaborative Perception

Open paper AskPDF arXiv

Collaborative Perception (CP) improves autonomous systems' awareness of their surroundings by sharing sensor data, intermediate features, and detection results. In real-world deployments, however, collaborating vehicles often use heterogeneous sensors, perception models, datasets, and training domains, creating feature-space shifts that degrade downstream fusion and detection. Existing approaches typically retrain fusion and detection components or introduce modality-specific feature interpreters. These methods scale poorly to newly joining agents and often require access to proprietary metadata, raising privacy concerns. We propose HeteroPROMPT, a real-time and privacy-preserving framework for heterogeneous collaborative perception. HeteroPROMPT rapidly aligns each heterogeneous agent's features with an ego-centric unified feature space through modular prompts and lightweight learning-based tuning, while keeping agent encoders and the collaborative fusion and detection stacks frozen. Its visual prompt-based training and inference modulate Bird's Eye View (BEV) features across channels and spatial locations with low computational overhead. For metadata-free deployment, an autoencoder learns a compact unified representation and extracts modality cues from shared features, enabling real-time modality classification and routing to the appropriate HeteroPROMPT modules without exposing proprietary agent information. Experiments on the OPV2V-H and V2XSet datasets show that HeteroPROMPT improves Average Precision over state-of-the-art heterogeneous CP methods while using orders of magnitude fewer trainable parameters. This offers a scalable and practical CP solution. The proposed modality classifier also predicts the joining agent's modality from compact features with greater than 99.99 percent accuracy during deployment. Code will be available at https://github.com/arminmaleki007/HeteroPROMPT.

Abstract

Explore similar work

HeteroPROMPT: A Real-time and Privacy-Preserving Heterogeneous Collaborative Perception Framework

GT-Space: Enhancing Heterogeneous Collaborative Perception with Ground Truth Feature Space

Faster-HEAL: An Efficient and Privacy-Preserving Collaborative Perception Framework for Heterogeneous Autonomous Vehicles