MotionKey Scribe

The idea is to teach a private motion vocabulary and turn an accepted gesture into a drawing. The interface and simulator let the recording, recognition, and command stages be examined separately.

Uncertainty has a visible outcome: no command. A recognized label still needs a fresh prediction, a valid arm lease, and a bounded glyph before the simulator accepts it.

The ML service records immutable data and compares engineered features, DTW, and a compact CNN. Evaluation separates recording sessions; the external wrist-IMU probe also holds people out and evaluates continuous background motion.

The gateway consumes persisted predictions rather than trusting client-supplied confidence. Pure geometry derives cable paths, while tests exercise replay protection, lease and freshness failures, canvas margins, step-rate bounds, and concurrent commands.

The documented external evaluation found an unsatisfactory tradeoff: the feature model detected more gestures but triggered too often, while the CNN largely abstained and missed most events. Neither result establishes useful recognition on the intended phone domain.

The documented smoke run connects a model prediction to deterministic simulated geometry. Firmware compilation establishes toolchain compatibility. Neither proves mechanical accuracy, physical stop behavior, serial transport, or a working phone-to-robot loop; no physical drawing result is claimed.