Selected work

Representative Papers

From Prompts to Pavement Through Time: Temporal Grounding in Agentic Scene-to-Plan Reasoning

May 19, 2026

Current large model–based approaches to autonomous driving scene understanding and planning lack effective temporal modeling, leading to inconsistent reasoning over sequential actions and compromising both safety and interpretability. To address this, this work proposes three multi-agent planner architectures incorporating varying degrees of temporal conditioning constraints. The authors establish the first empirical benchmark for temporally aware scene-to-planning reasoning on a subset of BDD-X and introduce evaluation metrics assessing semantic, syntactic, and logical consistency. Experimental results show that while explicit temporal constraints do not significantly improve standard NLP metrics, qualitative analysis reveals their capacity to elicit forward-looking risk assessment, stabilize corrective behaviors, and enhance strategic diversity. The study also highlights limitations in current prompt engineering practices regarding temporal grounding.

0 citationsRead paper

AROLA: A Modular Layered Architecture for Scaled Autonomous Racing

Feb 02, 2026

This work proposes AROLA, a modular and layered architecture for autonomous racing built upon ROS 2, addressing the limitations of existing fragmented or monolithic systems that lack standardized interfaces and hinder rapid component swapping and objective performance evaluation. AROLA decouples the autonomous driving pipeline into standardized functional layers—including perception, localization, planning, and control—and integrates a lightweight Race Monitor framework to enable real-time data logging and standardized post-race analysis. By enforcing uniform interfaces, the architecture supports plug-and-play module integration and facilitates reproducible benchmarking, significantly enhancing development efficiency and experimental comparability. The proposed system has been validated on both the RoboRacer simulation and physical platforms and was successfully deployed in the RoboRacer IV25 competition in 2025.

0 citationsRead paper

From Prompts to Pavement: LMMs-based Agentic Behavior-Tree Generation Framework for Autonomous Vehicles

Jan 18, 2026

This work addresses the limitations of traditional behavior trees in autonomous driving, whose static structures and reliance on manual tuning hinder adaptability in complex, dynamic environments—particularly impeding progress toward Level 5 autonomy. To overcome this, the authors propose an agent framework that integrates a large language model (LLM) with a multimodal vision model (LVM). When the baseline behavior tree fails, the system dynamically generates executable behavior subtrees through chain-of-symbol prompting and in-context learning, enabling adaptive decision-making without human intervention. This approach represents the first integration of chain-of-symbol prompting with behavior tree generation. Extensive experiments on the CARLA+Nav2 simulation platform demonstrate its effectiveness and generalization capability in handling unforeseen scenarios such as street blockages.

0 citationsRead paper

Domain Borders Are There to Be Crossed With Federated Few-Shot Adaptation

Jul 14, 2025

Federated learning (FL) faces three key challenges in resource-constrained edge environments: high annotation cost in target domains, significant covariate shift across clients, and communication- and energy-limited frequent model updates. To address low-shot target-domain adaptation, this paper proposes FedAcross+, a lightweight framework that freezes the pre-trained backbone and classifier while optimizing only a compact, learnable adapter layer. It supports streaming data processing and sporadic model updates, enabling robust adaptation to non-stationary edge environments. Crucially, FedAcross+ operates in a fully unsupervised manner on the target domain—requiring only a few (even single) unlabeled target samples for effective domain adaptation. Extensive experiments demonstrate that FedAcross+ significantly mitigates domain shift on low-resource edge devices, improves generalization, and achieves favorable trade-offs among communication efficiency, computational overhead, and deployment sustainability—establishing a practical new paradigm for edge-aware federated domain adaptation.

0 citationsRead paper

Recent publications

Latest Papers

From Prompts to Pavement Through Time: Temporal Grounding in Agentic Scene-to-Plan Reasoning

May 19, 2026

Current large model–based approaches to autonomous driving scene understanding and planning lack effective temporal modeling, leading to inconsistent reasoning over sequential actions and compromising both safety and interpretability. To address this, this work proposes three multi-agent planner architectures incorporating varying degrees of temporal conditioning constraints. The authors establish the first empirical benchmark for temporally aware scene-to-planning reasoning on a subset of BDD-X and introduce evaluation metrics assessing semantic, syntactic, and logical consistency. Experimental results show that while explicit temporal constraints do not significantly improve standard NLP metrics, qualitative analysis reveals their capacity to elicit forward-looking risk assessment, stabilize corrective behaviors, and enhance strategic diversity. The study also highlights limitations in current prompt engineering practices regarding temporal grounding.

0 citationsRead paper

AROLA: A Modular Layered Architecture for Scaled Autonomous Racing

Feb 02, 2026

This work proposes AROLA, a modular and layered architecture for autonomous racing built upon ROS 2, addressing the limitations of existing fragmented or monolithic systems that lack standardized interfaces and hinder rapid component swapping and objective performance evaluation. AROLA decouples the autonomous driving pipeline into standardized functional layers—including perception, localization, planning, and control—and integrates a lightweight Race Monitor framework to enable real-time data logging and standardized post-race analysis. By enforcing uniform interfaces, the architecture supports plug-and-play module integration and facilitates reproducible benchmarking, significantly enhancing development efficiency and experimental comparability. The proposed system has been validated on both the RoboRacer simulation and physical platforms and was successfully deployed in the RoboRacer IV25 competition in 2025.

0 citationsRead paper

From Prompts to Pavement: LMMs-based Agentic Behavior-Tree Generation Framework for Autonomous Vehicles

Jan 18, 2026

This work addresses the limitations of traditional behavior trees in autonomous driving, whose static structures and reliance on manual tuning hinder adaptability in complex, dynamic environments—particularly impeding progress toward Level 5 autonomy. To overcome this, the authors propose an agent framework that integrates a large language model (LLM) with a multimodal vision model (LVM). When the baseline behavior tree fails, the system dynamically generates executable behavior subtrees through chain-of-symbol prompting and in-context learning, enabling adaptive decision-making without human intervention. This approach represents the first integration of chain-of-symbol prompting with behavior tree generation. Extensive experiments on the CARLA+Nav2 simulation platform demonstrate its effectiveness and generalization capability in handling unforeseen scenarios such as street blockages.

0 citationsRead paper

Domain Borders Are There to Be Crossed With Federated Few-Shot Adaptation

Jul 14, 2025

Federated learning (FL) faces three key challenges in resource-constrained edge environments: high annotation cost in target domains, significant covariate shift across clients, and communication- and energy-limited frequent model updates. To address low-shot target-domain adaptation, this paper proposes FedAcross+, a lightweight framework that freezes the pre-trained backbone and classifier while optimizing only a compact, learnable adapter layer. It supports streaming data processing and sporadic model updates, enabling robust adaptation to non-stationary edge environments. Crucially, FedAcross+ operates in a fully unsupervised manner on the target domain—requiring only a few (even single) unlabeled target samples for effective domain adaptation. Extensive experiments demonstrate that FedAcross+ significantly mitigates domain shift on low-resource edge devices, improves generalization, and achieves favorable trade-offs among communication efficiency, computational overhead, and deployment sustainability—establishing a practical new paradigm for edge-aware federated domain adaptation.

0 citationsRead paper