Recent advances in large-scale multimodal learning have demonstrated remarkable capabilities in perception, reasoning, and instruction following. However, today's models remain limited by hallucinations, weak grounding, poor interpretability, and an inability to reliably connect perception with downstream decision-making. Addressing these limitations requires new algorithmic and representational foundations that integrate perception, reasoning, memory, and action while maintaining robustness and transparency.

The ultimate goal of my research is to develop trustworthy, interactive, and scalable multimodal AI systems that integrate vision, language, and action to enable grounded reasoning, predictive world modeling, and reliable decision-making in real-world environments. To this end, I develop methods spanning multimodal foundation models, computer vision, machine learning, robotics, reinforcement learning, and explainable AI to build autonomous systems that can perceive, understand, reason, and act reliably.

My earlier work focused on driving behavior analysis and intelligent transportation systems, where I developed deep learning and inverse reinforcement learning approaches for multimodal motion forecasting, scalable driver representation learning, and graph-based models for multi-agent interaction. These efforts established the foundations for my current research on grounded multimodal reasoning, trustworthy AI, world models, and embodied intelligence.

For the full list of my publications, visit my Google scholar page.