LLM-Assisted Coalition Formation for Cooperative Perception in Autonomous Driving

LLM-Assisted Coalition Formation for Cooperative Perception in Autonomous Driving Ahmad Sarlak, Hao Wang, Rahul Amin, Abolfazl Razi

arXiv:2608.00690v1 [cs.NI] 1 Aug 2026

School of Computing, Clemson University, Clemson, SC, USA Lincoln Laboratory, Massachusetts Institute of Technology, Lexington, MA, USA asarlak@g.clemson.edu, hao9@g.clemson.edu, rahul.amin@ll.mit.edu, arazi@clemson.edu

Abstract—Cooperative perception (CP) enables connected autonomous vehicles (CAVs) to share complementary observations for safer navigation, but practical deployment is limited by bandwidth constraints, unreliable links, and redundant information exchange. Existing CP methods often assume predefined participants and merely focus on collective perception. Likewise, recent LLM-based cooperative driving frameworks facilitate multi-vehicle reasoning but do not regulate participation criteria to select more beneficial vehicles. To bridge this gap, we propose an LLM-assisted coalition formation framework that selects the most informative helper vehicles before LLM reasoning. The approach jointly optimizes perceptual diversity using a determinantal point process (DPP) over multimodal vehicle embeddings and communication-aware reliability. This leads to a joint coalition selection and power allocation problem, which we solve efficiently via a relaxed convex reformulation and an ADMM-based optimization strategy that decouples diversityaware selection from network-aware resource allocation. The selected coalition is then summarized and provided with an LLM reasoning module for efficient and less redundant multi-vehicle decision support. Experimental results show that our approach outperforms other baselines in overall coalition value, while maintaining high diversity and improved networking efficiency. The framework achieves a better balance between task performance and safety across OPV2V and V2V4Real datasets, demonstrating its effectiveness for cooperative autonomous driving with communication constraints. Index Terms—cooperative perception, DPP, autonomous driving, LLM-assisted systems

I. I NTRODUCTION Autonomous driving has advanced rapidly with deep neural perception models, multimodal representations, and large-scale driving datasets [1]. However, safe driving in complex traffic environments remains challenging when a single connected autonomous vehicle (CAV) faces occlusion, limited sensing range, and varying wireless conditions [2]. Cooperative perception can alleviate this issue by allowing nearby vehicles to share observations and improve scene awareness. Yet, in practical vehicular networks, communication resources are limited, links are unreliable, and transmitting redundant information from many vehicles may increase latency and energy consumption without significantly improving perception quality [3]. Existing work on cooperative autonomous driving has mainly focused on perception and feature fusion for tasks such as detection and tracking [4]. Although these methods improve visibility beyond the sensing range of an individual vehicle, they often assume that participating vehicles are al-

ready given or selected using fixed or heuristic strategies [5]. In realistic traffic scenarios, however, neighboring vehicles do not contribute equally useful information: some offer redundant viewpoints, while others provide complementary observations that are crucial for resolving occlusions. At the same time, recent LLM-based cooperative driving frameworks, such as V2V-LLM [6], have shown the potential of using LLMs for multi-vehicle reasoning and decision support. However, existing cooperative perception methods focus on perception and typically assume predefined participants, while LLMbased cooperative driving frameworks reason over already collected multi-vehicle information. As a result, neither line of work addresses communication-aware helper selection before reasoning [7]. To address this gap, we propose an LLMassisted coalition formation framework that selects the most informative helper vehicles before LLM-based reasoning under communication constraints. Instead of aggregating information from all neighboring vehicles, the proposed framework selects the most informative helper vehicles based on both data diversity and communication efficiency. Each vehicle encodes its local multimodal observations into compact feature embeddings, and candidate coalitions are evaluated through a utility function that jointly captures perceptual diversity and communication efficiency. To promote complementary and non-redundant observations, we employ a determinantal point process (DPP)-based objective over normalized feature embeddings. To ensure practical cooperation, we further incorporate reliability-aware networking metrics, including throughput, retransmission effects, transmission distance, and power consumption. This results in a constrained joint optimization problem over coalition selection and transmit power. The selected coalition is then structured and provided to the LLM for reasoning and decision-making. Since the resulting problem is mixed-integer and nonlinear, direct optimization becomes intractable as the number of candidate vehicles increases. We therefore develop a scalable solution based on variable relaxation, log-domain power transformation, and the alternating direction method of multipliers (ADMM), which separates diversity-aware vehicle selection from network-aware resource allocation. Our framework uses the LLM as a reasoning module operating on structured summaries from the selected coalition, including scene coverage, detected objects, and environmental awareness. The proposed method therefore bridges communication-aware coalition op-

timization and high-level LLM reasoning, enabling more efficient, reliable, and non-redundant cooperative autonomous driving. The rest of the paper is organized as follows. Section II reviews related work. Section III presents the system model. Section IV introduces the performance metrics and problem formulation. Section V describes the proposed ADMM-based optimization method. Section VI provides simulation results, followed by conclusions. II. R ELATED W ORK A. Cooperative Perception in Autonomous Driving Cooperative perception has progressed from early featuresharing and intermediate-fusion methods to more structured and communication-efficient multi-agent perception. Early work such as F-Cooper demonstrated feature-level sharing for connected vehicles [8], while V2VNet introduced message passing for joint perception and prediction [9]. Later methods such as V2X-ViT and CoBEVT improved multi-agent fusion with attention and transformer-based designs across vehicles and infrastructure [10], [11]. More recent studies have shifted toward communication efficiency: Where2comm transmits spatially critical regions, while bandwidth-efficient communication modeling learns communication partners and informative content under bandwidth constraints [12]. Confidence-V2X uses confidence-guided sparse communication and gating, and optimized cooperative perception under imperfect communication studies helper selection and resource allocation under non-ideal links [5]. Overall, this line of work has substantially improved cooperative perception quality and bandwidth–accuracy tradeoffs, but it remains centered primarily on perception performance and does not explicitly address communication-aware helper selection under resource constraints. B. LLM-based Autonomous Driving In parallel, LLM-based autonomous driving has emerged as a promising direction for scene understanding, reasoning, and planning. DriveGPT4 and LMDrive showed that large language models can support interpretable or instructionfollowing end-to-end driving [13], [14], while DriveLM introduced graph-structured visual question answering over perception, prediction, and planning [15]. More recent multimodal driving models such as EMMA and DriveGPT4-V2 further unify multiple driving tasks within a common reasoning framework [16], [17]. In cooperative settings, V2VLLM demonstrated that a centralized language model can reason over perception features shared by multiple connected vehicles, and V2V-LLM extended this direction with graph-ofthought reasoning for cooperative perception, prediction, and planning [6]. Collectively, these works highlight the growing strength of LLM-based driving systems, but most operate after multi-vehicle information has already been collected or fused, without considering which vehicles should participate under communication constraints.

Fig. 1: LLM-assisted coalition formation for cooperative driving. III. S YSTEM M ODEL Consider a set of candidate helper vehicles N = {1, . . . , N }. Each vehicle i ∈ N encodes its local multimodal observations into a d-dimensional feature vector fi ∈ Rd . The corresponding unit-norm embedding is fi , i ∈ N. (1) f̃i = ∥fi ∥ A coalition S ⊆ N of size k = |S| shares its embeddings with a central aggregator (e.g., ego vehicle or edge server) to support cooperative perception and downstream LLM reasoning. For S = {i1 , . . . , ik }, define   F̃S = f̃i1 , f̃i2 , . . . , f̃ik ∈ Rd×k . (2) Each vehicle i ∈ S transmits a payload of Li bits over a wireless link of achievable rate Ri using power Pi . The nominal airtime is Li Ti = . (3) Ri Let Ptx = [P1 , . . . , PN ]⊤ collect the transmit powers. The utility of coalition S under power allocation Ptx is defined as

v S, Ptx = wa a(S) + wb b S, Ptx , (4) where wa , wb ≥ 0 and wa + wb = 1 denote the relative importance of the two objectives, namely a(S) measures perceptual diversity, and b(S, Ptx ) networking efficiency, defined in equations Eq.(11) and Eq.(18) respectively. Transmit power variables are defined for all candidate vehicles, while only vehicles in the selected coalition S actively transmit. The transmit power is subject to the following constraints: 0 < Pi ≤ Pmax , X Pi ≤ Ptot ,

∀i ∈ N ,

(5) (6)

i∈S

where Pmax is the per-vehicle transmit-power limit and Ptot is the total power budget. The objective is to select a coalition that maximizes information diversity while efficiently utilizing communication resources.

IV. P ERFORMANCE M ETRICS AND P ROBLEM F ORMULATION We next define the metrics used to quantify coalition quality. A. Spectral Diversity We represent coalition membership by a binary selection vector γ ∈ {0, 1}N satisfying 1⊤ γ = k,

C. Problem Formulation Given the diversity metric a(S) and the networking metric b(S, Ptx ), the joint coalition selection and power allocation problem is
tx v S(γ), P (19) max tx γ, P

1⊤ γ = k,

s.t.

(7)

where 1 is the all-ones vector. The corresponding coalition is S(γ) = {i ∈ N : γi = 1}. Collect all normalized embeddings in   F̃ = f̃1 , f̃2 , . . . , f̃N ∈ Rd×N .

(9)

(10)

(11)

This objective favors coalitions whose embeddings span a large volume in the embedding space, and thus capture complementary rather than redundant observations. The corresponding size-k diversity maximization problem is max

γ∈{0,1}N

s.t.

a(S(γ))

(12)

1⊤ γ = k.

(13)

B. Networking Efficiency To account for communication quality, we model the reliability of the link from vehicle i to the aggregator as   dηi ᾱi (Pi ) = exp −c , (14) Pi where di is the transmission distance, η is the path-loss exponent, and c > 0 absorbs the effects of noise and interference. The corresponding error probability is ϵi (Pi ) = 1 − ᾱi (Pi ). The effective throughput of vehicle i is
ζi (Pi ) = Ri 1 − ϵi (Pi ) ,

(15)

(16)

Under geometric retransmissions (i.e., repeated transmissions until success), the expected transmission energy is Pi Li Ei (Pi ) = , (17) 1 − ϵi (Pi ) We combine the two into a coalition-level networking score:
1 X (wT ζi (Pi ) − wE Ei (Pi )) , (18) b S, Ptx = |S| i∈S

where wT , wE ≥ 0 and wT + wE = 1. A coalition achieves a high networking score if it achieves high throughput with low energy consumption.

∀i ∈ N ,

γi Pi ≤ Ptot ,

(21) (22)

i=1

γi ∈ {0, 1},

We measure coalition diversity using the DPP-inspired score a(S) ≜ log det(KS,S ) .

N X

(8)

For a given coalition S, let F̃S denote the submatrix of F̃ indexed by S where each column corresponds to a selected vehicle in S. and define the Gram matrix k×k KS,S = F̃⊤ . S F̃S ∈ R

(20)

0 < Pi ≤ Pmax ,

∀i ∈ N .

(23)

Problem (19)–(23) is a mixed-integer nonlinear program (MINLP). The binary vector γ controls coalition composition and perceptual diversity, while Ptx determines the communication efficiency of the selected coalition. Direct solution is intractable for large candidate sets; therefore, we next develop a scalable ADMM-based solver. V. ADMM-BASED C OALITION O PTIMIZATION To enable scalable optimization, we reformulate the problem as a continuous optimization problem and solve it using ADMM. To obtain a tractable approximation of (19)–(23), we relax the binary variables to 0 ≤ γi ≤ 1,

1⊤ γ = k.

(24)

Under this relaxation, the diversity term a(S(γ)) is interpreted via a continuous log-determinant surrogate in γ. We also introduce the log-power variables yi = log Pi ,

Pi = eyi ,

(25)

which convert the power constraints into yi ≤ log Pmax , N X

∀i,

γi eyi ≤ Ptot .

(26) (27)

i=1

Under this transformation, the DPP term admits a standard continuous log-determinant relaxation in γ, while the negative networking objective admits a convex approximation in y under the adopted link model. Let b(γ, y) ≜

N N wE X wT X γi ζi (eyi ) − γi Ei (eyi ). k i=1 k i=1

(28)

The relaxed problem can then be written as min −wa a(S(γ)) − wb b(γ, y)

(29)

subject to (26), (27), and 0 ≤ γi ≤ 1, 1⊤ γ = k. Define the sets n o Γ = γ ∈ RN : 0 ≤ γi ≤ 1, 1⊤ γ = k ,

(30)

γ, y

n Y = y ∈ RN : yi ≤ log Pmax ,

N X i=1

o γi eyi ≤ Ptot .

(31)

Let IA (·) denote the indicator function of a set A. We split the objective into f (γ) = −wa a(S(γ)) + IΓ (γ),

(32)

g(γ, y) = −wb b(γ, y) + IY(γ) (y).

(33)

To enable variable splitting, we introduce two copies of the decision vector, x = (γ x , yx ) and z = (γ z , yz ), we obtain min f (γ x ) + g(γ z , yz )

(34)

x − z = 0.

(35)

x,z

s.t.

Using the scaled dual variable u, the augmented Lagrangian is ρ ρ Lρ (x, z, u) = f (γ x ) + g(γ z , yz ) + ∥x − z + u∥22 − ∥u∥22 . 2 2 (36) The ADMM iterations are i h ρ (37) xt+1 = arg min f (γ x ) + ∥x − z t + ut ∥22 , x 2 h i ρ z t+1 = arg min g(γ z , yz ) + ∥xt+1 − z + ut ∥22 , (38) z 2 ut+1 = ut + xt+1 − z t+1 . (39) The x- and z-updates are solved by projected gradient descent (PGD). Let Fx (x) and Fz (z) denote the objectives in (37) and (38), respectively. For each variable q ∈ {x, z}, the PGD update is 1

q ℓ+ 2 = q ℓ − α∇Fq (q ℓ ),   1 q ℓ+1 = ΠCq q ℓ+ 2 ,

(40) (41)

where ΠCq (·) denotes Euclidean projection onto the feasible set. For the x-update, Cx = Γ × RN , so the projection reduces to projecting γ onto the capped simplex Γ while leaving y unchanged. For the z-update, Cz = RN × Y(γ) , and the projection enforces the log-power box and sum-power constraints. The inner PGD loop terminates after a fixed number of iterations Tinner or when ∥q ℓ+1 − q ℓ ∥2 ≤ εpgd . ∥q ℓ ∥2

(42)

After ADMM converges, the transmit powers are recovered as Pi = eyi . The final discrete coalition is obtained by selecting the k largest entries of γ. VI. S IMULATION We evaluate the proposed framework in two settings: 1) synthetic large-scale coalition selection and 2) cooperative driving based on end-to-end LLM on the OPV2V [18] and V2V4Real [19] datasets. The synthetic studies examine optimization behavior and the trade-off between diversity and networking efficiency under controlled conditions, while the real-world study evaluates the impact of coalition optimization on downstream LLM reasoning. In the synthetic experiments, we consider a vehicular network with 20 candidate vehicles. Each vehicle is represented by a normalized feature embedding

Method Proposed k-DPP KMeans Random Channel-Aware

v(S) ↑ 1.65 ± 0.22 1.42 ± 0.24 1.04 ± 0.27 0.90 ± 0.29 1.03 ± 0.32

a(S) ↑ 0.84 ± 0.24 1.14 ± 0.07 0.67 ± 0.21 0.55 ± 0.26 0.60 ± 0.23

b(S) ↑ 0.32 ± 0.04 0.15 ± 0.07 0.15 ± 0.06 0.14 ± 0.06 0.17 ± 0.07

TABLE I: Coalition selection under equal power allocation.

of its local observation, and each communication link is characterized by distance-dependent reliability, channel rate, effective throughput, and energy cost. The proposed method jointly selects a coalition and allocates transmit power subject to coalition-size and transmit power constraints. Unless otherwise stated, all synthetic results are averaged over 100 independent random scenarios. Table I compares the method with k-DPP, KMeans, Random, and Channel-Aware (which selects vehicles based on channel quality) under equal power allocation (i.e., identical transmit power for all selected vehicles), ensuring a fair comparison focused solely on coalition selection. The proposed method achieves the highest overall coalition value v(S) and the best networking score b(S), while maintaining competitive diversity performance. Although k-DPP attains the highest diversity score a(S), its weaker networking efficiency reduces its overall utility. In contrast, our approach provides a more balanced trade-off, achieving diversity close to kDPP while significantly improving networking performance, leading to the best overall outcome. Other baselines, including KMeans, Random, and Channel-Aware Top-k, exhibit lower diversity and/or networking efficiency, resulting in inferior coalition values. These results demonstrate the importance of jointly considering diversity and communication efficiency, even under equal power allocation. Table II compares the method with several standard power allocation baselines, including equal power, random feasible allocation, channel inversion, and power-only projected gradient descent (PGD), under fixed coalition structures. The proposed method achieves the highest overall coalition value v(S) and networking score b(S), indicating more effective utilization of communication resources. The performance gain is primarily driven by a substantial reduction in energy consumption, while maintaining competitive throughput levels. In contrast, baseline methods achieve similar throughput but incur significantly higher energy consumption, which limits their overall utility. Among the baselines, power-only PGD improves upon heuristic approaches by reducing energy usage and increasing the networking score; however, it remains notably inferior to the our approach. The proposed approach achieves the lowest energy consumption by a wide margin, with only a modest reduction in throughput, resulting in the best trade-off between communication efficiency and reliability. These results demonstrate the effectiveness of the ADMMbased optimization in balancing throughput and energy under practical communication constraints. Fig. 2 shows the overall coalition value v(S) versus coali-

tion size k under fixed transmit power. The proposed method consistently achieves the highest utility across all coalition sizes. The performance gap widens with k, indicating more efficient utilization of cooperative resources. This demonstrates the advantage of the ADMM-based optimization in balancing diversity, communication efficiency, and resource allocation.

Fig. 2: Overall coalition value versus coalition size. We next consider a realistic vehicular communication setting with heterogeneous feature embeddings and nonidentical links. In this scenario, coalition selection must jointly balance perceptual diversity and networking efficiency. For comparison, baseline methods use simple heuristics: k-DPP selects vehicles based on diversity with random power allocation, highest-α selects vehicles with the most reliable links, and closest-k selects the nearest vehicles based on distance. As shown in Fig. 3, the proposed method consistently achieves the highest diversity score across coalition sizes.

selected helper summaries are provided to Llama3, served locally via Ollama with deterministic decoding (T = 0), to perform structured multi-vehicle reasoning. The LLM input consists of structured summaries derived from multi-vehicle observations, including regional object counts and distancebased spatial features. The model predicts multiple perception labels along with a final driving decision, which is further refined using a rule-based risk score. Ground-truth labels are generated from oracle summaries constructed using the full five-vehicle fused view, ensuring a consistent and fullinformation reference for evaluation. We evaluate both decision quality and safety-related behavior using a comprehensive set of metrics. Specifically, We report different metrics. Mean Field Accuracy evaluates perlabel prediction accuracy, while Action Accuracy reflects the accuracy of the final driving decision. Unsafe Under-Reaction quantifies the frequency of insufficiently cautious responses in critical situations, whereas Over-Conservative measures unnecessary cautious actions. Additionally, Field Macro-F1 and Action Macro-F1 capture balanced performance across perception fields and action classes, respectively. Tables III and IV summarize the results on V2V4Real and OPV2V, respectively. Across both datasets, the proposed method consistently achieves the best overall performance, including higher decision accuracy and improved structured prediction quality. It also maintains competitive or improved safety-related metrics, reducing unsafe under-reaction while avoiding overly conservative behavior. These results demonstrate that communication-aware coalition selection improves both structured scene understanding and downstream decisionmaking. Overall, our approach provides a more effective balance between performance and safety compared to diversitybased and random selection strategies within the LLM-driven cooperative driving pipeline. VII. C ONCLUSION

(a)

Fig. 3: Diversity under realistic communication conditions. Finally, we evaluate the proposed framework in an endto-end LLM-based cooperative driving setting on both the OPV2V and V2V4Real datasets. For each frame, a candidate set of five helper vehicles is constructed, and two helpers are selected using different coalition selection strategies. The

This paper presented an LLM-assisted coalition formation framework for cooperative autonomous driving under communication constraints. Our approach addresses the gap between communication-aware vehicle selection and LLM-based multivehicle reasoning by selecting the most informative helper vehicles before reasoning. We formulated joint coalition selection and power allocation as an optimization problem that integrates DPP-based diversity with communication efficiency metrics, and solved it using a scalable ADMM-based approach. Results showed that the proposed method consistently outperforms baseline coalition selection strategies in overall utility, while achieving higher diversity and better networking efficiency. Experiments on OPV2V and V2V4Real datasets further demonstrated improved downstream cooperative driving performance, with a better balance between task accuracy, safety-related behavior, and communication efficiency. These findings suggest that optimized coalition formation is a practical front-end for LLM-based cooperative driving, enabling more reliable and less redundant information sharing in realistic vehicular networks.

v(S) ↑ 2.39 ± 1.63 2.11 ± 1.82 2.44 ± 1.68 2.76 ± 1.54 6.22 ± 1.20

Method Equal power Random feasible Channel inversion Power-only PGD Proposed

b(S) ↑ 0.15 ± 0.18 0.12 ± 0.20 0.16 ± 0.19 0.20 ± 0.17 0.59 ± 0.13

Throughput ↑ 2.158 ± 0.34 2.10 ± 0.36 2.15 ± 0.34 2.15 ± 0.34 2.00 ± 0.35

Energy ↓ 3.12 ± 0.63 3.18 ± 0.70 3.10 ± 0.71 2.90 ± 0.56 0.59 ± 0.14

TABLE II: Power allocation under fixed coalitions. Method Proposed Random k-DPP

Mean Field Acc (%) ↑ 56 54 54

Action Acc (%) ↑ 48 46 45

Unsafe Under-Reaction (%) ↓ 52 52 53

Over-Conservative (%) ↓ 10 9 9

TABLE III: Performance of LLM-based cooperative driving on the V2V4Real dataset under different coalition selection strategies. Metric Action Acc (%) ↑ Unsafe Under-Reaction (%) ↓ Field Macro-F1 (%) ↑ Action Macro-F1 (%) ↑

Proposed 10.77 89.28 25.12 8.36

k-DPP 8.90 91.00 24.00 6.23

Random 6.10 94.96 12.00 3.28

TABLE IV: Performance comparison of coalition selection strategies on the OPV2V dataset under different helper selection strategies.

R EFERENCES [1] A. Sarlak, A. Razi, X. Chen, and R. Amin, “Diversity maximized scheduling in roadside units for traffic monitoring applications,” in 2023 IEEE 48th Conference on Local Computer Networks (LCN). IEEE, 2023, pp. 1–4. [2] A. Sarlak and Y. Darmani, “An approach to improve the quality of service in dtn and non-dtn based vanet,” Journal of Information Systems and Telecommunication (JIST), vol. 4, no. 32, p. 240, 2021. [3] A. Sarlak, R. Amin, and A. Razi, “Extended visibility of autonomous vehicles via optimized cooperative perception under imperfect communication,” Transportation Research Part C: Emerging Technologies, vol. 180, p. 105350, 2025. [4] Q. Chen, S. Tang, Q. Yang, and S. Fu, “Cooper: Cooperative perception for connected autonomous vehicles based on 3d point clouds,” in 2019 IEEE 39th International Conference on distributed computing systems (ICDCS). IEEE, 2019, pp. 514–524. [5] X. Tan, R. Wang, J. Wang, S. Wang, X. Wang, and D. Wu, “Confidencev2x: Confidence-driven sparse communication for efficient v2x cooperative perception,” Advanced Engineering Informatics, vol. 69, p. 103914, 2026. [6] H.-k. Chiu, R. Hachiuma, C.-Y. Wang, S. F. Smith, Y.-C. F. Wang, and M.-H. Chen, “V2v-llm: Vehicle-to-vehicle cooperative autonomous driving with multi-modal large language models,” arXiv preprint arXiv:2502.09980, 2025. [7] A. Sarlak and A. Razi, “Reliability-aware determinantal point processes for robust informative data selection in large language models,” arXiv preprint arXiv:2602.00885, 2026. [8] Q. Chen, X. Ma, S. Tang, J. Guo, Q. Yang, and S. Fu, “F-cooper: Feature based cooperative perception for autonomous vehicle edge computing system using 3d point clouds,” in Proceedings of the 4th ACM/IEEE Symposium on Edge Computing, 2019, pp. 88–100. [9] T.-H. Wang, S. Manivasagam, M. Liang, B. Yang, W. Zeng, and R. Urtasun, “V2vnet: Vehicle-to-vehicle communication for joint perception and prediction,” in European conference on computer vision. Springer, 2020, pp. 605–621. [10] R. Xu, Z. Tu, H. Xiang, W. Shao, B. Zhou, and J. Ma, “Cobevt: Cooperative bird’s eye view semantic segmentation with sparse transformers,” arXiv preprint arXiv:2207.02202, 2022.

[11] R. Xu, H. Xiang, Z. Tu, X. Xia, M.-H. Yang, and J. Ma, “V2x-vit: Vehicle-to-everything cooperative perception with vision transformer,” in European conference on computer vision. Springer, 2022, pp. 107– 124. [12] Y. Hu, S. Fang, Z. Lei, Y. Zhong, and S. Chen, “Where2comm: Communication-efficient collaborative perception via spatial confidence maps,” Advances in neural information processing systems, vol. 35, pp. 4874–4886, 2022. [13] Z. Xu, Y. Zhang, E. Xie, Z. Zhao, Y. Guo, K.-Y. K. Wong, Z. Li, and H. Zhao, “Drivegpt4: Interpretable end-to-end autonomous driving via large language model,” IEEE Robotics and Automation Letters, vol. 9, no. 10, pp. 8186–8193, 2024. [14] H. Shao, Y. Hu, L. Wang, G. Song, S. L. Waslander, Y. Liu, and H. Li, “Lmdrive: Closed-loop end-to-end driving with large language models,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024, pp. 15 120–15 130. [15] C. Sima, K. Renz, K. Chitta, L. Chen, H. Zhang, C. Xie, J. Beißwenger, P. Luo, A. Geiger, and H. Li, “Drivelm: Driving with graph visual question answering,” in European conference on computer vision. Springer, 2024, pp. 256–274. [16] Z. Xu, Y. Bai, Y. Zhang, Z. Li, F. Xia, K.-Y. K. Wong, J. Wang, and H. Zhao, “Drivegpt4-v2: Harnessing large language model capabilities for enhanced closed-loop autonomous driving,” in Proceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 17 261– 17 270. [17] J.-J. Hwang, R. Xu, H. Lin, W.-C. Hung, J. Ji, K. Choi, D. Huang, T. He, P. Covington, B. Sapp et al., “Emma: End-to-end multimodal model for autonomous driving,” arXiv preprint arXiv:2410.23262, 2024. [18] R. Xu, H. Xiang, X. Xia, X. Han, J. Li, and J. Ma, “Opv2v: An open benchmark dataset and fusion pipeline for perception with vehicle-tovehicle communication,” in 2022 International Conference on Robotics and Automation (ICRA). IEEE, 2022, pp. 2583–2589. [19] R. Xu, X. Xia, J. Li, H. Li, S. Zhang, Z. Tu, Z. Meng, H. Xiang, X. Dong, R. Song et al., “V2v4real: A real-world large-scale dataset for vehicle-to-vehicle cooperative perception,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 13 712–13 722.

Related documents

• Perception-Aware Communication Middleware for Distributed Visual Perception in UAV Swarms#1051358
• One Token Per Frame: Reconsidering Visual Bandwidth in World Models for VLA Policy#1129249
• VeriFuse: Bounded Vision-Language Arbitration and Reason-Guided Refinement for Cooperative 3D Perception#1035730
• When Agents Disagree: The Selection Bottleneck in Multi-Agent LLM Pipelines#973096
• Symbolic Guidance for LLM Agents in Distributed Multiagent Coordination#1115919