I am a senior staff engineer at Qualcomm working on physical AI and autonomous driving. My research interests lie at the intersection of physical AI and generative AI. Specifically I am interested in multimodal large language models, world models, and their application for end-to-end autonomous driving.

Prior to joining Qualcomm, I did my PhD in Electrical Engineering at SUTD under supervision of Chau Yuen where I received the best PhD dissertation award for my PhD thesis. I also spent a year at Linköping University as a post doc working with Erik Larsson.

Selected Awards and Grants

  • Qualcomm's IP Achievement Award, 2025 (extremely competitive).
  • Best PhD Thesis Award from Singapore University of Technology and Design, 2019.
  • FIRST Industry Workshop Outstanding Graduate Research Award, 2017.
  • MediaTek Graduate Research Competition Award, 2017.
  • Merlion PhD Award, 2014.
  • President Graduate Fellowship from Singapore University of Technology and Design, 2013.
  • Singapore Graduate Fellowship Award, 2013.

Selected Publications

MAPLE paper thumbnail

MAPLE: Latent Multi-Agent Play for End-to-End Autonomous Driving

What if a self-driving model could rehearse against its own traffic — and learn from every near-miss? We introduce MAPLE: closed-loop, multi-agent training for VLA driving models, played out entirely in latent space — no simulator required.
Core Idea: Reactive Rollouts 🤝 RL
The formula: 🚗 + 🚙🚕 → 🔁 (latent play) → 🏆 SOTA on Bench2Drive

Generative Scenario Rollouts paper thumbnail

Generative Scenario Rollouts for End-to-End Autonomous Driving

What if your autonomous agent had an "internal theater" to rehearse the future? We introduce GeRo: a VLA model augmented with a latent world model for autonomous driving.
Core Idea: World Model 🤝 VLA
The formula: 👁️ + 📝 → N × (🧠 + 🔮 → 🎮)

Adversarial Attacks on Radio Signal Classification paper thumbnail

Adversarial Attacks on Deep-Learning Based Radio Signal Classification

As the pioneering work that introduced adversarial attacks to the wireless physical layer, this research exposes a critical vulnerability in DL-based systems. We demonstrate that nearly invisible adversarial perturbations can systematically trick modulation classifiers, outperforming traditional jamming techniques and fundamentally challenging the security of AI-driven communications.