塚田研究室のメンバーは、2025年6月22日から25日までルーマニアのクルジュ＝ナポカで開催された第36回IEEE Intelligent Vehicle Symposium（IV2025）に参加しました。この著名な国際会議には、知能化車両分野の第一線の研究者や実務家が一堂に会しました。本研究室からは塚田准教授、ジャヴァンマルディ特任助教、博士課程の学生であるYun LiさんとYuze Jiangさんが参加しました。また、昨年塚田研究室に客員学生として滞在していたMikel Garcíaさんとも現地で再会することができました。
私たちのチームは、2件の論文発表と2件の基調講演を通じてシンポジウムに大きく貢献しました。博士課程のLiさんは、大規模言語モデルを用いて自動運転システムを特定の運転嗜好に適合させる新しいフレームワーク「PrefDrive: Enhancing Autonomous Driving through Preference-Guided Large Language Models」を発表しました。同じく博士課程のJiangさんは、物体検出を改善するためにLiDARセンサーの配置を最適化する新手法を提案する「Towards Efficient Roadside LiDAR Deployment: A Fast Surrogate Metric Based on Entropy-Guided Visibility」に関する研究を紹介しました。
学生の発表に加え、塚田准教授は第2回セキュアコネクテッドビークルに関するワークショップで「V2X Communication Technologies in the Era of End-to-End Autonomous Driving」と題した基調講演を行いました。また、ジャヴァンマルディ特任助教は第14回協調型自動運転に関するワークショップで「From Lab to Road: Advances and Challenges in V2X Cooperative Perception for AVs」と題した基調講演を行いました。これらの発表は、自動運転と高度交通システムの未来を前進させるという当研究室の継続的な取り組みを示すものです。
発表
Li, Yun, Javanmardi, Ehsan, Thompson, Simon, Katsumata, Kai, Orsholits, Alex, Tsukada, Manabu, "PrefDrive: Enhancing Autonomous Driving through Preference-Guided Large Language Models", In: 36th IEEE Intelligent Vehicles Symposium (IV2025), Cluj-Napoca, Romania, 2025.Proceedings Article | Abstract | Links | BibTeX
@inproceedings{Li2025c,
title = {PrefDrive: Enhancing Autonomous Driving through Preference-Guided Large Language Models},
author = {Yun Li and Ehsan Javanmardi and Simon Thompson and Kai Katsumata and Alex Orsholits and Manabu Tsukada},
url = {https://github.com/LiYun0607/PrefDrive/
https://huggingface.co/liyun0607/PrefDrive
https://huggingface.co/datasets/liyun0607/PrefDrive},
doi = {10.1109/IV64158.2025.11097672},
year = {2025},
date = {2025-06-22},
urldate = {2025-06-22},
booktitle = {36th IEEE Intelligent Vehicles Symposium (IV2025)},
address = {Cluj-Napoca, Romania},
abstract = {This paper presents PrefDrive, a novel framework that integrates driving preferences into autonomous driving models through large language models (LLMs). While recent advances in LLMs have shown promise in autonomous driving, existing approaches often struggle to align with specific driving behaviors (e.g., maintaining safe distances, smooth acceleration patterns) and operational requirements (e.g., traffic rule compliance, route adherence). We address this challenge by developing a preference learning framework that combines multimodal perception with natural language understanding. Our approach leverages Direct Preference Optimization (DPO) to fine-tune LLMs efficiently on consumer-grade hardware, making advanced autonomous driving research more accessible to the broader research community. We introduce a comprehensive dataset of 74,040 sequences, carefully annotated with driving preferences and driving decisions, which, along with our trained model checkpoints, will be made publicly available to facilitate future research. Through extensive experiments in the CARLA simulator, we demonstrate that our preference-guided approach significantly improves driving performance across multiple metrics, including distance maintenance and trajectory smoothness. Results show up to 28.1% reduction in traffic rule violations and 8.5% improvement in navigation task completion while maintaining appropriate distances from obstacles. The framework demonstrates robust performance across different urban environments, showcasing the effectiveness of preference learning in autonomous driving applications. },
keywords = {},
pubstate = {published},
tppubtype = {inproceedings}
}
This paper presents PrefDrive, a novel framework that integrates driving preferences into autonomous driving models through large language models (LLMs). While recent advances in LLMs have shown promise in autonomous driving, existing approaches often struggle to align with specific driving behaviors (e.g., maintaining safe distances, smooth acceleration patterns) and operational requirements (e.g., traffic rule compliance, route adherence). We address this challenge by developing a preference learning framework that combines multimodal perception with natural language understanding. Our approach leverages Direct Preference Optimization (DPO) to fine-tune LLMs efficiently on consumer-grade hardware, making advanced autonomous driving research more accessible to the broader research community. We introduce a comprehensive dataset of 74,040 sequences, carefully annotated with driving preferences and driving decisions, which, along with our trained model checkpoints, will be made publicly available to facilitate future research. Through extensive experiments in the CARLA simulator, we demonstrate that our preference-guided approach significantly improves driving performance across multiple metrics, including distance maintenance and trajectory smoothness. Results show up to 28.1% reduction in traffic rule violations and 8.5% improvement in navigation task completion while maintaining appropriate distances from obstacles. The framework demonstrates robust performance across different urban environments, showcasing the effectiveness of preference learning in autonomous driving applications.
Jiang, Yuze, Javanmardi, Ehsan, Tsukada, Manabu, Esaki, Hiroshi, "Towards Efficient Roadside LiDAR Deployment: A Fast Surrogate Metric Based on Entropy-Guided Visibility", In: 36th IEEE Intelligent Vehicles Symposium (IV2025), Cluj-Napoca, Romania, 2025.Proceedings Article | Abstract | Links | BibTeX
@inproceedings{Jiang2025,
title = {Towards Efficient Roadside LiDAR Deployment: A Fast Surrogate Metric Based on Entropy-Guided Visibility},
author = {Yuze Jiang and Ehsan Javanmardi and Manabu Tsukada and Hiroshi Esaki},
url = {https://arxiv.org/abs/2504.06772},
doi = {10.1109/IV64158.2025.11097554},
year = {2025},
date = {2025-06-22},
urldate = {2025-06-22},
booktitle = {36th IEEE Intelligent Vehicles Symposium (IV2025)},
address = {Cluj-Napoca, Romania},
abstract = {The deployment of roadside LiDAR sensors plays a crucial
role in the development of Cooperative Intelligent
Transport Systems (C-ITS). However, the high cost of LiDAR
sensors necessitates efficient placement strategies to
maximize detection performance. Traditional roadside LiDAR
deployment methods rely on expert insight, making them
time-consuming. Automating this process, however, demands
extensive computation, as it requires not only visibility
evaluation but also assessing detection performance across
different LiDAR placements. To address this challenge, we
propose a fast surrogate metric, the Entropy-Guided
Visibility Score (EGVS), based on information gain to
evaluate object detection performance in roadside LiDAR
configurations. EGVS leverages Traffic Probabilistic
Occupancy Grids (TPOG) to prioritize critical areas and
employs entropy-based calculations to quantify the
information captured by LiDAR beams. This eliminates the
need for direct detection performance evaluation, which
typically requires extensive labeling and computational
resources. By integrating EGVS into the optimization
process, we significantly accelerate the search for optimal
LiDAR configurations. Experimental results using the AWSIM
simulator demonstrate that EGVS strongly correlates with
Average Precision (AP) scores and effectively predicts
object detection performance. This approach offers a
computationally efficient solution for roadside LiDAR
deployment, facilitating scalable smart infrastructure
development. },
keywords = {},
pubstate = {published},
tppubtype = {inproceedings}
}
The deployment of roadside LiDAR sensors plays a crucial
role in the development of Cooperative Intelligent
Transport Systems (C-ITS). However, the high cost of LiDAR
sensors necessitates efficient placement strategies to
maximize detection performance. Traditional roadside LiDAR
deployment methods rely on expert insight, making them
time-consuming. Automating this process, however, demands
extensive computation, as it requires not only visibility
evaluation but also assessing detection performance across
different LiDAR placements. To address this challenge, we
propose a fast surrogate metric, the Entropy-Guided
Visibility Score (EGVS), based on information gain to
evaluate object detection performance in roadside LiDAR
configurations. EGVS leverages Traffic Probabilistic
Occupancy Grids (TPOG) to prioritize critical areas and
employs entropy-based calculations to quantify the
information captured by LiDAR beams. This eliminates the
need for direct detection performance evaluation, which
typically requires extensive labeling and computational
resources. By integrating EGVS into the optimization
process, we significantly accelerate the search for optimal
LiDAR configurations. Experimental results using the AWSIM
simulator demonstrate that EGVS strongly correlates with
Average Precision (AP) scores and effectively predicts
object detection performance. This approach offers a
computationally efficient solution for roadside LiDAR
deployment, facilitating scalable smart infrastructure
development.
role in the development of Cooperative Intelligent
Transport Systems (C-ITS). However, the high cost of LiDAR
sensors necessitates efficient placement strategies to
maximize detection performance. Traditional roadside LiDAR
deployment methods rely on expert insight, making them
time-consuming. Automating this process, however, demands
extensive computation, as it requires not only visibility
evaluation but also assessing detection performance across
different LiDAR placements. To address this challenge, we
propose a fast surrogate metric, the Entropy-Guided
Visibility Score (EGVS), based on information gain to
evaluate object detection performance in roadside LiDAR
configurations. EGVS leverages Traffic Probabilistic
Occupancy Grids (TPOG) to prioritize critical areas and
employs entropy-based calculations to quantify the
information captured by LiDAR beams. This eliminates the
need for direct detection performance evaluation, which
typically requires extensive labeling and computational
resources. By integrating EGVS into the optimization
process, we significantly accelerate the search for optimal
LiDAR configurations. Experimental results using the AWSIM
simulator demonstrate that EGVS strongly correlates with
Average Precision (AP) scores and effectively predicts
object detection performance. This approach offers a
computationally efficient solution for roadside LiDAR
deployment, facilitating scalable smart infrastructure
development.
Tsukada, Manabu, "V2X Communication Technologies in the Era of End-to-End Autonomous Driving", Keynote talk at The 2nd Workshop on Secure connected vehicles: Digital Twin, UAVs, and Smart Transportation, at IEEE IV 2025, Cluj-Napoca, Romania, 22.06.2025.Presentation | Abstract | Links | BibTeX
@misc{Tsukada2025b,
title = {V2X Communication Technologies in the Era of End-to-End Autonomous Driving},
author = {Manabu Tsukada},
url = {https://sites.google.com/view/b-stem-iot/},
year = {2025},
date = {2025-06-22},
urldate = {2025-06-22},
address = {Cluj-Napoca, Romania},
abstract = {Autonomous driving technology is undergoing a significant paradigm shift from traditional rule-based systems to integrated End-to-End (E2E) deep learning architectures. This transition necessitates a fundamental rethinking of Vehicle-to-Everything (V2X) communication, as existing V2X standards, primarily designed for rule-based systems, may not fully leverage the capabilities or address the needs of E2E models. This presentation explores the evolution required for V2X technologies in the E2E era. We contrast rule-based and E2E architectures, highlighting the limitations of current V2X approaches like object-level message sharing for E2E systems that benefit from richer data. While intermediate feature sharing via V2X is promising, its practical implementation faces hurdles, notably the heterogeneity of sensors, AI models, and tasks across vehicles. To address these challenges, we introduce a research approach aiming to maximize V2X value through an E2E pipeline encompassing data foundation (Co3SOP dataset for collaborative 3D semantic occupancy), perception adaptation (PHCP framework for heterogeneous collaboration during inference), and decision optimization (PrefDrive integrating LLMs with preference learning). Through these interconnected efforts, we aim to unlock the full potential of V2X communication to enhance the safety, efficiency, and robustness of E2E autonomous driving systems.},
howpublished = {Keynote talk at The 2nd Workshop on Secure connected vehicles: Digital Twin, UAVs, and Smart Transportation, at IEEE IV 2025},
keywords = {},
pubstate = {published},
tppubtype = {presentation}
}
Autonomous driving technology is undergoing a significant paradigm shift from traditional rule-based systems to integrated End-to-End (E2E) deep learning architectures. This transition necessitates a fundamental rethinking of Vehicle-to-Everything (V2X) communication, as existing V2X standards, primarily designed for rule-based systems, may not fully leverage the capabilities or address the needs of E2E models. This presentation explores the evolution required for V2X technologies in the E2E era. We contrast rule-based and E2E architectures, highlighting the limitations of current V2X approaches like object-level message sharing for E2E systems that benefit from richer data. While intermediate feature sharing via V2X is promising, its practical implementation faces hurdles, notably the heterogeneity of sensors, AI models, and tasks across vehicles. To address these challenges, we introduce a research approach aiming to maximize V2X value through an E2E pipeline encompassing data foundation (Co3SOP dataset for collaborative 3D semantic occupancy), perception adaptation (PHCP framework for heterogeneous collaboration during inference), and decision optimization (PrefDrive integrating LLMs with preference learning). Through these interconnected efforts, we aim to unlock the full potential of V2X communication to enhance the safety, efficiency, and robustness of E2E autonomous driving systems.
Javanmardi, Ehsan, "From Lab to Road: Advances and Challenges in V2X Cooperative Perception for AVs", Keynote talk at 14th Workshop and Industry Panel on Cooperative Automated Driving and Future Mobility Systems at IEEE IV2025, Cluj - Napoca, Romania, 22.06.2025.Presentation | Links | BibTeX
@misc{Javanmardi2025,
title = {From Lab to Road: Advances and Challenges in V2X Cooperative Perception for AVs},
author = {Ehsan Javanmardi},
url = {https://www.ika.rwth-aachen.de/en/ieee-iv-workshop-25.html},
year = {2025},
date = {2025-06-22},
urldate = {2025-06-22},
address = {Cluj - Napoca, Romania},
howpublished = {Keynote talk at 14th Workshop and Industry Panel on Cooperative Automated Driving and Future Mobility Systems at IEEE IV2025},
keywords = {},
pubstate = {published},
tppubtype = {presentation}
}
