Research

Machine Learning & AI, Predictive Science, Decision Science, Connected and Autonomous Vehicles, High Performance Computing, Data Science.

Selected Projects

Predictive trajectory planning for autonomous vehicles at intersections using reinforcement learning

In this work we put forward a predictive trajectory planning framework to help autonomous vehicles plan future trajectories. We develop a partially observable Markov decision process (POMDP) to model this sequential decision making problem, and a deep reinforcement learning solution methodology to learn high-quality policies. The framework demonstrates promising performance for planning horizons as long as five seconds. We compare safety, comfort, and energy efficiency of the planned trajectories against human-driven trajectories in both experimental driving environments, and demonstrate that it outperforms human-driven trajectories in a statistically significant fashion in all aspects.
[PDF]

Real-time Driving Risk Prediction and Mitigation for Connected Vehicles: Real-time Risk Prediction, Trajectory Forecasting, and Real-World Testbeds

(2019-2022, collaborate with DENSO International American Inc.)

In this long-term and impactful project, I propose a learning-based framework along with corresponding algorithms to solve cutting-edge problems for connected and autonomous vehicles. Detailed topics include trajectory prediction for traffic agents, real-time driving risk prediction and learning-based motion planning. The project started in 2019 with DENSO, which is a high-prestige company working on intelligent transportation field.

We have led efforts to publish three peer-reviewed manuscripts and two patents under this topic area. We also tested our algorithms in real-world CV transportation systems in Ann Arbor, MI, and Dublin, OH.

Earlier Research (2017-2019)

Increasing GPS Localization Accuracy with Reinforcement Learning

RL-based algorithm performance vs. EKF

In this work, I propose a reinforcement learning framework to increase GPS localization accuracy. The framework does not make rigid assumptions on the GPS device hardware parameters or motion models, nor does it require infrastructure-based reference locations. It learns an optimal strategy to make “corrections” on raw GPS observations. The model uses an efficient confidence-based reward mechanism, which is independent of geolocation, thereby enabling the model to be generalized. We incorporate a map matching-based regularization term to reduce the variance of the reward return. The reinforcement learning model is constructed using the asynchronous advantage actor-critic (A3C) algorithm. The proposed reinforcement learning model converges fast, has less prediction variance, and can localize vehicles with 50% less error compared to the benchmark Extended Kalman Filter model.

[PDF]

V2XSim: A V2X Simulator for Connected and Automated Vehicle Environment Simulation

In this work we propose a vehicle-to-everything (V2X) simulator, called V2XSim, for connected vehicle environment simulation. The work provides an integrated V2X simulation platform, built with the Gazebo robot simulation engine. The simulated world is constructed with two types of models, namely dynamic and static models. We provide a detailed communication architecture to simulate the vehicle communication process. In this architecture, a closed-loop control module is built to pass control commands and control vehicle models without human intervention. We also provide multiple APIs for users to simplify the vehicle control process and make the simulator easy to use.

[Github, PDF]

GPU-Based Parallel Computing Algorithm for Real-time Mapping Between Large-scale Networks

In this work, I propose a scalable massively-parallel algorithm to solve the general mapping problem in large-scale networks in real-time. The proposed parallel algorithm takes advantage of GPU architecture and launches millions of workers to calculate values on a target network simultaneously. Threads are managed through the SIMT execution model and target values are updated through atomic operations. Experiments show the proposed algorithm can accomplish network mapping (find importance weights for links in a real-world large-scale shared-mobility network) with more than 2 million weights within 1.82 μs. Compared to serial algorithms, the speedup is more than 12,000 times.

[PDF]