autonomous driving/trajectory prediction/accident anticipation/world models/connected vehicles
I am a Ph.D. student at Purdue University, advised by Prof. Yiheng Feng. My research focuses on safe and interpretable autonomous driving: forecasting how surrounding road users will move, anticipating accidents before they happen, and grounding language and vision in world models that let a vehicle reason about what comes next.
Before Purdue, I was a research assistant at the State Key Laboratory of Internet of Things for Smart City, University of Macau, working with Prof. Zhenning Li and Prof. Chengzhong Xu.
I like ideas that cross disciplines — several of my models borrow from how human drivers perceive and reason — and the engineering it takes to make them work outside the lab. I have built connected-vehicle systems end to end and deployed them on real vehicles, and I enjoy untangling complex systems problems along the way.
From simulation to real roads, and the tools I build on the way.
A roadside system that predicts which vehicles are about to cut into a highway work zone: a 4 s time-to-collision screen, then an attention-LSTM that forecasts the next 4 s of each flagged vehicle. Field-tested at work zones in Indiana; ≈ 11 ms per prediction.
Read the project →Getting the warning into the car: a digital twin of the work zone feeds a roadside unit, which broadcasts over C-V2X to an onboard unit and an in-vehicle display I built — now installed in a test vehicle.
Read the project →A dashboard pane for context, cost, rate limits, cache, tools and subagents, with reply themes to match. Twelve styles, each redrawing the whole pane in its own visual language — pick one below.
Big subagent batches turn into a radial run trace that grows in real time.
aurora
Vision-only V2X perception with LLM-assisted semantic localization. Team lead.
LLaMA-7B-Chat fine-tuned on ~600k tourism texts for multi-turn travel Q&A.
UI mock-ups from text with Stable Diffusion and multimodal LLMs.
* equal contribution · Google Scholar
H. Liao*, Y. Li*, Z. Li, C. Wang, Y. Guan, K. H. Tam, C. Tian, L. Li, C. Xu
A three-stage framework that couples object detection with LLM-based scene prompting to predict when, where and what kind of accident will happen — 14.6% higher precision, and accidents anticipated 16.4% earlier.
H. Liao*, Y. Li*, Z. Li, C. Wang, C. Tian, Y. Huang, Z. Bian, K. Zhu, G. Li, Z. Pu, et al.
A lightweight trajectory predictor modeled on how human drivers distribute their attention, distilled from a larger teacher so that it stays small without giving up accuracy.
H. Liao, H. Shen, B. Wang, Y. Li, Y. Tang, C. Wang, D. Zhuang, K. Chen, et al.
B. Hu, Z. Lu, H. Liao, C. Yuan, B. Rao, Y. Li, G. Li, Z. Cui, C. Xu, Z. Li
H. Liao*, Y. Li*, Z. Li, C. Wang, Y. Guan, K. H. Tam, C. Tian, L. Li, C. XuCo-first
H. Liao*, Y. Li*, Z. Li, C. Wang, C. Tian, Y. Huang, Z. Bian, K. Zhu, G. Li, Z. Pu, et al.Co-first
H. Liao*, Y. Li*, Z. Li, C. Wang, Z. Cui, S. E. Li, C. XuCo-first
H. Liao*, Y. Li*, Z. Li, Z. Bian, J. Lee, Z. Cui, G. Zhang, C. XuCo-first
H. Liao, S. Liu, Y. Li, Z. Li, C. Wang, B. Wang, Y. Guan, C. Xu
H. Liao, C. Wang, Z. Li, Y. Li, B. Wang, G. Li, C. Xu
Bilingual notes on the building blocks of deep learning. All posts →
Why sequences need weights that remember, and how backpropagation runs through time.
Read → Positional embeddingWhere the sines and cosines come from, what they tell a Transformer about order, and the code.
Read → Loss functionsThree PyTorch losses that look alike, and when each one is the right call.
Read →Ideas, collaborations, compute, or questions about my work — my inbox is open.