Research Focus

I view the generalization failure of current robot foundation models as a context uncertainty problem: when the current observation is insufficient for action selection, a robot must maintain, retrieve, and translate task-relevant context into guidance that a low-level policy can execute. Vision-language-action (VLA) models inherit broad visual and semantic knowledge from pretrained VLMs, but this knowledge is not always grounded in a form that supports reliable action generation after robot-data fine-tuning.

To address this, I study hierarchical VLM–VLA systems, in which a VLM-based high-level policy keeps track of task progress, episodic memory, and embodiment-specific context, reasons about the uncertainty of the current situation, and provides structured guidance to a low-level VLA policy. My work has progressed from real-world VLA deployment, through capacity allocation for VLA fine-tuning and action-level subtask discovery, to explicit language memory for non-Markovian manipulation. Going forward, I am interested in agentic VLMs that decide what to remember, what to retrieve, and how to guide the VLA, choosing among text memory, keyframe retrieval, visual traces, keypoints, or latent guidance as the situation demands.