>
AyakaCorp

Prototype · live demo · Apache-2.0

ayaka-hand

A hand-prosthesis control system that pre-shapes the grip before contact. A wrist camera looks at the object being reached for and forms the grasp on the way in; the muscle signal is reduced to a single binary GO.

The user decides when to close. The vision system decides how.

status
prototype · public demo
license
Apache-2.0
student
2,232,839 params
weights
8.9 MB ONNX
runtime
in-browser · ORT Web
grasp budget
≤125 ms
01 · The problem conventional control: pick the grip, meter the force, consciously, every time

The problem

≈ 23 % of upper-limb myoelectric prostheses are abandoned in real use.

A recurring reason is cognitive load. Conventional control asks the user to consciously modulate muscle contractions to pick a grip and meter force, for every object, all day. ayaka-hand moves that work to a vision system and leaves the user exactly one decision.

the user decides when.
the vision system decides how.
02 · Two tiers the teacher is never in the real-time loop

How it works

Tier 1 · teacher · offline

Cortex

model
Qwen3-VL-32B-Instruct
license
Apache-2.0
in the loop
never

A vision-language model that reasons about an object's appearance and state and emits a structured grasp label. It labels the training set offline; in the demo it appears only as a slower second-opinion panel.

Tier 2 · reflex · on-device

Student

model
MobileNetV2 · grip + force heads
size
2,232,839 params · 8.9 MB
input
wrist-camera pixels only

Distilled from the teacher's labels, this is what runs in real time, in the browser via ONNX-Runtime Web. No object detector tells it what it sees.

A grasp must be committed inside ≤125 ms to feel like part of the body. A cloud model cannot live in that budget, so it doesn't.

Two-tier architecture: the Cortex teacher labels data offline; the distilled Student model runs on-device in real time. The user's EMG signal remains a single binary GO trigger. OFFLINE · TEACHER · NEVER IN THE REAL-TIME LOOP COCO 2017 crops 11,004 train · 613 val Cortex · teacher (VLM) Qwen3-VL-32B-Instruct · Apache-2.0 structured grasp labels · distillation ON-DEVICE · REAL TIME GRASP COMMITTED IN ≤125 MS Wrist camera pixels only Student · reflex (CNN) MobileNetV2 · 2,232,839 params grip + force heads · 8.9 MB ONNX decides how grip force Pre-shaped grip formed on approach GO close EMG · user one binary trigger decides when
Two tiers. The teacher labels data offline; the distilled student runs on-device. EMG stays one binary GO.
Two-tier architecture: the Cortex teacher labels data offline; the distilled Student model runs on-device in real time. The user's EMG signal remains a single binary GO trigger. OFFLINE · NEVER IN THE LOOP COCO 2017 crops 11,004 train · 613 val Cortex · teacher (VLM) Qwen3-VL-32B-Instruct distilled labels ON-DEVICE · ≤125 MS Wrist camera pixels only Student · reflex (CNN) MobileNetV2 · 2,232,839 params grip + force heads · 8.9 MB ONNX grip + force Pre-shaped grip formed on approach GO close EMG · user · one binary trigger decides when · vision decides how
Two tiers. The teacher labels data offline; the distilled student runs on-device. EMG stays one binary GO.
03 · Results 11,004 train · 613 val crops the baseline is handed the true class name; the student sees pixels

Results

The baseline is deliberately generous: it gets the ground-truth object class name (a perfect detector) and looks up a canonical grip for it. The student gets only pixels.

Grip and force metrics, lookup baseline versus student model
Metric Lookup baseline
gets the true class name
Student
2.23M · pixels only
Grip accuracy 70.5 % 72.3 %
Grip macro-F1 0.581 0.659
Force accuracy 65.4 % 81.1 %
Force macro-F1 0.595 0.806

Grip accuracy

Lookup baseline70.5 %
Student72.3 %

Grip macro-F1

Lookup baseline0.581
Student0.659

Force accuracy

Lookup baseline65.4 %
Student81.1 %

Force macro-F1

Lookup baseline0.595
Student0.806

Pinch grip · macro-F1

0.000 0.323

No object class maps cleanly to a pinch; you have to look at the object. The baseline cannot score here. The student can.

Force accuracy · gain

+15.7 pts

Force depends on an object's state (a full cup versus an empty one), which a class name cannot encode.

04 · Boundaries AMD Developer Hackathon · Act II

What it is not

  • 1A research prototype from a hackathon submission. Not a medical device. No injury-prevention claims are made.
  • 2No physical hand has been built. The demo renders a simulated hand in the browser.
  • 3The teacher can emit advisory hazard flags (hot, sharp, fragile). They are display-only: not distilled into the student, not wired to actuation. The system does not detect or prevent harm and must not be relied on to.
05 · Data & licensing COCO 2017 · CC BY / CC BY-SA

Data and licensing

Trained on 11,004 teacher-labelled COCO 2017 crops, validated on 613. Images were license-filtered to CC BY, CC BY-SA and no-known-restriction sources; everything NonCommercial or NoDerivatives was excluded, because the crops are derivative works. Code and documentation are Apache-2.0.

06 · Links open in a new tab

Links

Open the live demo ↗ lfl1337.github.io/ayaka-hand
View the repository ↗ github.com/lfl1337/ayaka-hand