Skip to content

HLM-Micro-gesture-v0

6-class gesture recognition from 6-axis accelerometer + gyroscope data in the HLM-Micro tier. Same architecture as the other Micro checkpoints — only the classifier head and training task differ.

Access

Trained weights and inference code are available through the Early Access program or a commercial engagement.

Task

6-class gesture classification on an IMU window (~1 second at typical wearable sampling rates).

ClassSignature (synthetic proxy)
stillDevice at rest
shakeRapid lateral oscillation
tapSharp single impulse
circleSmooth yaw-dominant rotation
walkQuasi-periodic gait rhythm
throwAccelerate → release → impact → decay

Headline

The public checkpoint is trained on synthetic gesture signatures and reaches ceiling accuracy on the training-time benchmark. Real-world IMU data (user variability, sensor noise, sampling drift) will land lower — expect 75–92% on real data. Commercial engagement includes retraining on your target user population.

Hardware integration

Designed to be wire-compatible with common ESP32 + MPU6050 firmware used in the Industrial edge hardware reference. A model trained on this task can be dropped into a physical sensor node without feature-layout changes.

Positioning

ScenarioFit
Consumer wearable gesture shortcutsPrimary — with per-user calibration
Industrial operator-action recognitionStrong — gloved-hand gesture control
Accessibility / assistive devicesStrong — minimal motion discrimination

Out-of-scope

  • Not trained on real human-motion data. The public checkpoint uses synthetic gesture signatures. Production retraining on UCI HAR / MobiAct / a proprietary dataset is the next step.
  • No per-user calibration. Deployable wearables typically train a small adaptation layer per user on first wear.
  • Gesture vocabulary is illustrative, not exhaustive. Production deployments typically define their own vocabulary.

How to get started

Early Access or commercial — same path as the other HLM-Micro checkpoints.

License

BSL 1.1 (Apache 2.0 in 2030) once Early Access is granted.