Yunpeng(Larry) Liu

I'm a fourth-year PhD student at University of British Columbia, where I'm advised by Frank Wood. I'm also an Mitacs intern at Invereted AI, where we are working on advancing autonomous driving with generative models. I interned with the Model Alignment team at Autodesk AI Lab, where I developed reinforcement learning post-training pipelines for Qwen VLM on agentic, structured visual editing tasks. Previously, I completed my Bachelor of Applied Science in Electrical and Computer Engineering at the University of British Columbia.

My current research is on learning from third-person demonstrations with video diffusion models. Previously, I worked on modelling realistic multi-agent behaviors.

Email  /  Scholar  /  LinkedIn  /  Github

Refereed Publications

We show that synthetic data from post-trained diffusion-based video world models can support imitation learning with scaling behavior similar to ground-truth demonstrations.

Diffusion models based closed-loop autoregressive multi-agent traffic scene simulation.

Improving online continual learning performance by optimizing the experience replay loss in a way that stabilizes neural network representations between parameter updates.

A Diffusion model based method of generating traffic scenarios (DJINN). we demonstrate how DJINN flexibly enables direct test-time sampling from a variety of valuable conditional distributions including goal-based sampling, behavior-class sampling, and scenario editing.

Existing multi-agent behavioral models rely on manually annotated HD maps which is expensive and time-consuming. We propose an aerial image-based map representation that requires minimal annotation and provides rich road context information for traffic agents like pedestrians and vehicles.

Combines RL value function estimation and Bayesian inference to infer dynamically feasible trajectories given low-level controllers trained for each vehicle type.