Mingyi Wang

Now I'm a first-year Ph.D. student at Arizona State University, supervised by Prof. Shaofeng Zou. Previously, I was a Master's student in the College of Control Science and Engineering at Zhejiang University, supervised by Prof. Guang Li. Prior to that, I obtained my B.Eng. from Zhejiang University.

Email  /  CV  /  Scholar  /  Github

Research

I am interested in Reinforcement Learning, Robot Navigation, Autonomous Driving.

An evaluation framework for transferring LLM modules to autonomous driving motion generation. Tokenizer, positional embedding, training, post-training, and test-time computation are systematically assessed on the Waymo Sim Agents benchmark, with effective adaptations identified for improved motion generation performance.

A hybrid SSI based on the converter between bimodal electrophysiological signals and audio signals. EEG and sEMG are fused through two sequence-to-sequence models, and multi-task losses are applied to achieve complementarity between speech intention and muscle activity in silent speech.

A simple yet effective joint prediction methodology for multi-agent trajectory generation in autonomous driving applications, enhancing the accuracy of agent interaction modeling and reducing the collision rate among agents.

A discrete States-based Trajectory Planning (DSTP) algorithm for autonomous nonholonomic robots, improving trajectory smoothness, efficiency, and control through optimized multi-variable representation and L-BFGS optimization.

An improved quadratic programming approach for path planning in autonomous driving that optimizes path smoothness and curvature using the Frenét frame, enhancing comfort and safety in urban environments.

A lidar-camera fusion system for terrain classification in off-road robots, leveraging multiple optical channels in Lidars and depth-enhanced images to classify various terrain types.

Experience

After graduating from my undergraduate studies, I spent my first year as an intern at my advisor's company, primarily working on spherical robot trajectory navigation using ROS and C++.

From August 2023 to February 2024, I interned at DJI Automotive in Shenzhen. Under the guidance of my mentor, Lu Zhang, I conducted research on learning-based trajectory generation methods and a unified framework for prediction and planning. Ultimately, this project was integrated into DJI Automotive's current end-to-end model.

In April 2024, I became a visiting student under Professor Kaicheng Yu at Westlake University. I then joined UDEER.AI for an internship, where I worked on a joint project exploring advanced agent interaction modeling methods and experimented with integrating them with reinforcement learning. This approach aimed to address the challenge of distribution shift in learning-based trajectory planning.

All these experiences have led me to where I am today.

While visiting Westlake University as a student, I joined the Planning, Decision, and Control group at UDEER.AI to participate in an industry-academic collaborative project. Here, I worked on cutting-edge robotics decision-making algorithms and completed the work titled Mitigating Out-of-Distribution Issues in Motion Prediction with Autoregressive Model and Reinforcement Learning.

My internship experience at DJI Automotive was incredibly rewarding. Our team proposed the PnP model, which was eventually applied to autonomous vehicles. After returning from the internship, I was inspired by my mentor Lu Zhang's paper at the company, and I completed my scene-consistent trajectory prediction paper: A Joint Prediction Method of Multi-Agent to Reduce Collision Rate.

Personal page under continuous development...