Research
My overall goal is to enable robots to perform and learn manipulation tasks using insights from human motor learning. I first model the hierarchical structure of human manipulation—from high-level brain planning to low-level muscular execution—and then transfer those representations and learning mechanisms to robots.
Cross-modal brain–muscle modulation
I align EEG and EMG signals in a shared representation space through proxy cross-modal generation and motor classification. The resulting models make abstract cortical–muscular control measurable and provide interpretable accounts consistent with cognitive neuroscience findings.
Multi-scale motor modeling
To reflect the hierarchy of manipulation, I use complementary modalities at multiple scales: video for sequence-level analysis, EEG for primitive-level decoding, and EMG for action-level segmentation and rhythm modeling.
Online skill decoding
My skill-decoding pipeline combines source-imaging-based cross-electrode alignment, task-driven signal denoising, and adaptive data- and representation-level alignment. Online updates support rapid calibration despite limited, noisy, and variable biosignal data.
Multi-module robot learning
Inspired by the complementary roles of the cerebellum, basal ganglia, and cortex, I study collaborative imitation, reward-driven policy optimization, and memory-driven self-learning for sample-efficient transfer and zero-shot skill acquisition.
Stage 1 · Human manipulation modeling
Human manipulation has a hierarchical organization: high-level planning in the brain and low-level execution through the muscular system. My work makes this implicit sensorimotor structure explicit through cross-modal brain–muscle representations and multi-scale video, EEG, and EMG models. Results have appeared in IEEE Transactions on Cybernetics, IEEE Transactions on Medical Robotics and Bionics, IEEE Transactions on Emerging Topics in Computational Intelligence, IEEE Transactions on Industrial Informatics, ACM Multimedia, and IJCNN.
Stage 2 · Human–machine skill transfer
Transferring human skills to robots requires robust decoding from limited, low-SNR data and learning systems that can improve from sparse demonstrations. I combine online skill decoding with collaborative imitation, policy optimization, and generative self-refinement to form a practical basis for bidirectional human–robot learning.