ayaka-hand: the grip forms before you touch

Roughly 23% of upper-limb myoelectric prostheses are abandoned in real-world use. It's rarely the hardware that fails. The real cost is cognitive load: consciously picking a grip and metering force for every object, all day long. ayaka-hand moves that decision off the user and onto a wrist camera. The camera looks at what you're reaching for and pre-shapes the hand into the right grip before contact. Your own muscle signal, just one EMG channel, gets reduced to a single binary GO. You decide when to close. The vision system already decided how. Underneath, it's a teacher and a reflex. The teacher is a cloud vision-language model, Qwen3-VL. It labels training data offline, and in this live demo it also shows up as a slower second opinion a couple of seconds later; it's never in the real-time control loop. The reflex is what actually decides: a 2.23-million-parameter model running on-device, in the browser, on pixels alone, committing within 125 milliseconds. It never gets told an object's class name, only what it looks like. On force estimation specifically, that matters: it beats a baseline that's simply handed the object's exact label for free, and still wins by 16 points of accuracy, because force depends on an object's state, not its label. This is a research prototype, not a certified medical device, and what's real today is the live demo. There's no physical hand built yet. Hazard flags from the cloud model are advisory only; they're never wired to an actuator. The teacher's serving path targets AMD ROCm, and I validated it end-to-end on RDNA3 hardware, with MI300X as the scale target; the student was trained on local hardware. I'm building it because I want to see it actually help someone use their hand again, not because I wanted to win a hackathon. Try the live demo yourself in any browser: https://lfl1337.github.io/ayaka-hand