Xiaoyu Liang

This project studies how vision-language-action (VLA) models can be restructured to run efficiently on neural processing units, using smolVLA as a concrete case. It focuses on the hardware–model–compiler interface: how vision backbones, transformer layers, and action heads must be decomposed into kernels and dataflows that match AI Engine constraints, and what this reveals about performance bottlenecks and scalability for embodied AI at the edge.

This project builds a large-scale, clinician-grounded dataset capturing physical caregiving with synchronized tactile, visual, pose, gaze, and annotation streams. Its aim is to provide the empirical substrate for robots to learn safe, comfortable assistance strategies from expert demonstrations, enabling research on multimodal fusion, contact-rich interaction, and clinically meaningful human–robot collaboration.

This project develops whole-arm soft tactile skins that allow robots to treat contact as a rich feedback channel rather than a hazard. It examines how deformations of compliant surfaces translate into interpretable capacitive signals under noise, wear, and motion, with the broader goal of enabling safe, human-aligned manipulation in close physical proximity.

This project explores paint-based mutual capacitive sensing as a low-cost way to turn arbitrary robot geometries into large-area proximity and touch sensors. It analyzes how conductive coatings, electrode layouts, and signal-processing strategies jointly determine sensing range, resolution, and scalability, aiming to make rich surface perception accessible for industrial and collaborative robots through simple fabrication processes.