AI turns text into electric body feedback with 63% accuracy

Galvanic vestibular stimulation (GVS) — electrical currents applied behind the ears to modulate balance and motion perception — has long been used in research, but generating specific waveforms that match intended sensations remained a challenge. A new paper from the University of Tsukuba tackles this by combining a custom dataset with a generative AI model. The team collected 100 distinct GVS waveforms and 1,526 free-form sensation descriptions from 16 participants, revealing that waveforms consistently evoked fewer semantic categories (8.18 vs. 9.45) and a higher dominant-category proportion (26.97% vs. 21.25%) than random permutations — meaning the electrical signals carry predictable perceptual signatures.

To turn text into stimulation, they built a retrieval-guided one-dimensional convolutional variational autoencoder. When 10 new participants were tested on matching waveforms to visual cues, they correctly identified congruent pairs 63.33% of the time (d'=0.70, p<0.001), significantly above chance. This proof-of-concept suggests that GVS could become a programmable modality for embodied feedback — think VR haptics without wearables, or assistive cues for balance-impaired users — all controlled by natural language descriptions of the desired sensation.