Research Directions </>

LLM-Powered V2X Intelligence for CAV & ITS

LLM-driven understanding, prediction, and advisory across connected corridors and digital twins.

V2XSynth: LLM + RAG for Realistic V2X Scenario Synthesis
NEW

An LLM-orchestrated, retrieval-augmented framework that compiles natural-language intents into fully parameterized CARLA simulations. A dual Interpreter/Agent performs intent decomposition and grounded retrieval across behavior, map, and V2X (SPaT/MAP/BSM) profiles, assembling Scenic+Python with automated spatial anchoring (signal richness, network centrality). Executions inject realistic communication degradations (latency, jitter, loss, decode delay) and collect telemetry for closed-loop refinement—enabling structured exploration of V2X-induced faults and safety evaluation.

V2X-LLM: LLM-Powered V2X Integration and Understanding
NEW

I developed a V2X-LLM framework to enhance Vehicle-to-Everything (V2X) communication by integrating advanced large language models (LLMs). The architecture consists of a data pipeline that encodes these inputs into structured prompts, enabling the LLM to perform advanced reasoning tasks such as scenario explanation, V2X data description, state prediction, and navigation advisory generation. Additionally, I incorporated digital twin technology to enable real-time monitoring, simulation, and reasoning, further enhancing the system's capability to predict and optimize traffic flows and vehicle behaviors. This system improves data interpretation and contextual understanding within connected corridors, facilitating seamless interactions between vehicles, infrastructure, and other components.

Digital Twins & Simulation with V2X Data Pipelines

Real-time physical–virtual integration, standards-compliant data ingestion, and closed-loop control.

Digital Twin Framework for Physical-Virtual Integration
NEW

I developed a scalable Digital Twin architecture for a real-world C-V2X connected vehicle corridor, built upon an enhanced data pipeline that ingests and synchronizes SPaT, MAP, and BSM messages via distributed edge servers. It details modules for message retrieval, decoding, extraction with temporal synchronization and spatial matching, and storage. In the virtual domain, CARLA and Unreal Engine are used to reconstruct intersections, vehicles, RSUs, and traffic signals with real-time fidelity. The system supports bidirectional feedback—dynamic signal timing adjustments, vehicle advisory speed recommendations, and incident notifications—demonstrating improved data integration, simulation accuracy, and proactive traffic management.

C-V2X Connected Corridor Data Pipeline Development

I developed a data pipeline system using socket programming to connect with Roadside Unit (RSU) devices at intersections, enabling the acquisition of real-time traffic signal data. I processed this data in accordance with SAE J2735 standards and managed its storage using a PostgreSQL database. Additionally, I created an interactive dashboard to display both historical and real-time traffic signal data from RSUs. My work also involved conducting tests with On-Board Unit (OBU) devices in connected vehicles to receive and process information from RSU devices at intersections. Furthermore, I implemented a digital twin to simulate the collected data in a high-fidelity virtual environment, facilitating real-time monitoring, comprehensive analysis, and behavior simulation.

Autonomous Driving Simulation & Modeling

I conducted extensive research on collaborative automated driving systems (CADS) using the CARLA simulation environment. This involved simulating the operational life cycle of CADS (sensing - prediction - planning - control), and validating the algorithms developed for collaborative driving. My focus for the sensing function was on developing and integrating roadside camera-based object detection and tracking systems. For prediction, I embedded and verified the effectiveness of motion prediction algorithms within the environment. In terms of planning and control, I designed and explored vehicle control strategies and deep reinforcement learning models, executing path planning as well as longitudinal and lateral vehicle control.

Interaction-Aware Reasoning for Multi-Agent Motion & Active Safety

Relational models (graphs/hypergraphs) fused with vehicle dynamics for predictive safety and planning.

AI²-Active Safety: HGNN + Vehicle Dynamics for Probabilistic HF-TTC Early Warning
NEW

I developed AI²-Active Safety framework that fuses a physics-grounded bicycle model—augmented with road-gradient effects—and a hypergraph-based neural network to predict probabilistic, multi-agent trajectories. By numerically integrating host-vehicle dynamics via fourth-order Runge–Kutta and inferring group-wise ambient behavior through a transformer-based HGNN, the system computes stochastic high-fidelity Time-to-Collision (HF-TTC) distributions that capture both complex interactions and behavioral uncertainty. This approach significantly outperforms constant-velocity TTC and state-of-the-art baselines in early-warning accuracy while faithfully representing real-world crash rarity.

Relational Reasoning for Motion: Graph/Hypergraph-Based Interaction-Aware Planning

I have explored the dynamics of relational reasoning and interaction modeling, particularly for trajectory prediction and planning. This involved constructing sophisticated models capable of learning dynamic interaction graphs and hypergraphs based on behaviors of agents, with a particular emphasis on predicting future relational dynamics. I have integrated this knowledge into interaction-aware decision-making processes, combining it with multi-modal motion prediction and the modeling of physical dynamics to enhance motion planning and predictive capabilities.

Interaction-Aware Multi-Modal Trajectory Prediction with Future-Guided Embeddings

I developed a dynamic graph approach to accurately represent bidirectional and heterogeneous interactions within vehicle motion states. I also proposed a future-guided graph embedding approach to link historical and future motion states, highlighting the correlation of future motions between surrounding vehicles and the target vehicle. Additionally, I provided predictions for both longitudinal and lateral driving behaviors in a multi-modal manner, assigning probabilities to potential future trajectories.

Spatial-Temporal Modeling & Network Reconstruction

Sparse-sensor data fusion and imputation for robust transportation network analytics.

Spatial-Temporal Imputation: Reconstructing Transportation Data from Sparse Sensors

I proposed a generic, cost-effective informatization framework within an intelligent transportation system, significantly reducing reliance on fixed sensors by using mobile sensors for improved data collection and accessibility. I developed a generic sensor network that utilized sparsely located sensors, including fixed sensors, drones, and floating vehicles, to gather data with characterized spatial and temporal sparsity. Furthermore, I specialized in inferring and reconstructing unobserved information by training models with partially observed data, enabling the generalization of traffic data reconstruction in real-world scenarios with missing values.

↑