Infrastructure-based Autonomous Driving

Motion/Behavior Planning

A main area of our research is to develop safe and efficient methods of autonomous vehicle management in different road settings, currently with a focus on unsignalized intersections. To coordinate multiple autonomous vehicles at unsignalized intersections, we employ various methods based on model-based optimization and/or data-driven generative models. 

Collaborative Perception

In scenarios like urban intersections with a high density of vehicles and active interactions among them, understanding the overall road situation and predicting the paths of many vehicles are essential. 

However, current standalone vehicle perception technology faces challenges in navigating such environments. This is due to limitations caused by occlusion and restricted sensor measurement distances, making it difficult to perceive a comprehensive area accurately. 

Therefore, Collaborative Perception aims to actively utilize sensor data from high-positioned infrastructure to achieve precise detection and classification of various-sized objects across a broader area.


Learning-based Control

Safe Control with Uncertain Perception

In autonomous systems, we must plan a safe motion avoiding surrounding agents' predicted future motions. However, this is challenging because the predictions are uncertain and multimodal. We aim to plan a safe motion under uncertainty.

1) We design safe motion planning systems based on predictions for the surrounding agent's future motions.

2) We formulate a chance-constrained optimization problem to guarantee safety theoretically. We approach it as a sampling-based method and a Gaussian Mixture Model(GMM) based method.

3) We improve learning-based prediction models to enhance the prediction quality so that we can control agents more safely and effectively.

Safe Control under Distribution Shifts

1) Uncertainty Quantification

We aim to develop techniques that can effectively characterize and account for uncertainties in mathematical models used for control system design. By quantifying these uncertainties, engineers can design control systems that are robust to variations and disturbances, ensuring stable performance even in the presence of model inaccuracies or disturbances.

2) Anomaly Detection

We aim to detect incompleteness and anomalies in Deep Neural Network (DNN) models tailored for semiconductor manufacturing processes. In particular, we develop techniques for identifying and mitigating outliers that could adversely affect process control and yield. By improving the robustness and reliability of DNN models, this research endeavors to enhance process efficiency and product quality in the semiconductor industry.

Smart City Project

The Smart City Project aims to implement autonomous driving technology beyond Level 4. 

1) Reality-based System

The Smart City Project utilizes a scaled-down map of real-world intersections at a 1:15 ratio. Consequently, vehicles, infrastructure, and other structures are also reduced to a 1:15 scale, and we refer to this as a "semi-realistic" testbed. This feature distinguishes it from simulations by incorporating realistic sensors, communication, and control noise, providing a more realistic data collection and physical movement experience. Furthermore, the infrastructure, such as traffic lights, is considered "intelligent" with sensors and computing modules. Unlike devices designed for passive control of current traffic flow, this "intelligent" infrastructure actively uses continuous information sharing with vehicles through V2X technology. This approach is pivotal in future technology development, enabling active traffic flow control and cooperative perception. 

2) Implementation of Complex Traffic Flow

The intersection traffic island induces situations requiring vehicle coordination, such as merging and diverging movements. Additionally, aiming to create a dense traffic scenario within the intersection, autonomous vehicles contribute to maximizing vehicle density with a target production of up to 30 units. Moreover, the outer circular road is designed as a unidirectional loop, functioning like a highway, ensuring that autonomous vehicles can move to any desired destination from any location on the road. This setup guarantees an endlessly navigable system, evaluating the capabilities of autonomous driving in an urban autonomous driving service scenario.

3) High Suitability for Multi-Agent Technology Development

In experiments involving the control of multiple vehicles, collisions between vehicles, pedestrians, or other structures are inevitable considerations. Such accidents not only pose significant safety concerns but also incur substantial financial burdens. This project actively leverages its intermediate nature between reality and simulation to conduct experiments with various agents, ensuring minimal safety issues and comparatively lower financial burdens in the event of accidents. Furthermore, intentional staging of scenarios, especially edge cases, enables the creation and collection of data related to realistic situations such as accidents or lane departures.