Research project
Embodied intelligence and research agents
Task-state representations, long-horizon navigation, and autonomous experimentation.
I study how agents represent their progress, use feedback, and recover from mistakes over long horizons.
StageNav uses geometric verification and state-aligned reinforcement learning for long-horizon navigation. ForesightNav combines semantic mapping, graph-based target verification, and geometric exploration.
My internships at Baidu and ByteDance have extended this interest to research agents that manage context, plan experiments, and refine models. I am exploring scientific-state representations for autonomous PDE modeling and structured temporal-spatial modeling for ocean prediction; these manuscripts remain under review.
