Large-Scale 3DGS Sensor Simulation: Founded and scaled the 3D Gaussian Splatting (3DGS) sensor simulation at Waymo from a small project to over one million reconstructions. This initiative reconstructs realistic 3D scenarios from sensor logs for safety evaluations and has grown from a team of 3 to a 20+ person effort, shaping Waymo's next-gen ML-based simulation.

Sensor Diffusion: Contributed to a research project using video diffusion models to generate synthetic sensor data for testing autonomous vehicle behaviors, with work published in IROS 2025 (Oral). This evolved to use the world model, Genie3, for multi-modal (Camera, Lidar) simulation of rare events.

SceneDiffuser and SceneDiffuser++: Founded and developed the SceneDiffuser model, the core of SceneAI, for large-scale, controlled scenario creation and traffic generation, with research published in NeurIPS 2024. SceneDiffuser++ expanded this for city-scale traffic generation, with work published in CVPR 2025.

SceneAI: Directed the development of the generative AI engine for the team's data synthesis platform, delivering tools to generate traffic, densify rare events, and create scenarios from natural language.

Agentic Simulation Realism Triage: Proposed and developed a Gemini-based agent to triage the realism of simulation results, enhancing its tool-use and reasoning capabilities.

Agentic Scene Generation: Developed a Gemini-based agent using a diffusion model for end-to-end synthetic scenario generation from natural language and images. This project won a company-wide hackathon and was launched in the data synthesis platform.

AutoDense: Developed a diffusion model to generate rare events from log, supporting multiple launches.

Responsible for model design & implementation & training, model integration with production simulation system, cross team collaboration, production usage landing and onboard engineers engagement.

AITrafficGen: Landed our first generative AI product: AITrafficGen, which support generate realistic traffic in large scale(>100k scenarios) to discover long-tail events. Model produces 16 agents, with 12s trajectory of each, similar to diffusion based video generation. The synthetic set created with it already used in prod launches.

AutoDense: Proposed and building our second generative AI product for automatic long tail events densification, support constrainted diffusion to generate similar rare events given seed long tail events to densify evaluation signal. This is similar to image guided diffusion. Backbone of it is based on Perceiver IO, and used transformer as encoder & denoiser. We build the model as large multi-task foundational model.

LLM Guided SceneGen: Inspired by DALL·E 3, proposed third generative AI product use multi-modal LLM to guide our diffusion model generate synthetic scenario end-to-end given natural language description and image to boost the synthetic data generation at Waymo. Initial protoptype shows very promising result.