feat(sim): synthetic IMU sensor-data generator for fall/crash scenarios (domain-randomised)

[nandal]

Background

The acute fast-path detector (design discussion #58; epics #28 / #37) needs training and test data, and real labelled falls are scarce. Good news: a fall/crash is physically describable — rigid-body kinematics projected into the body frame yields accelerometer (≈ 2nd derivative + gravity) and gyro (angular velocity) traces. We already have a simulation scaffold in app/test/sim/.

The task

Build a generator that synthesises labelled sensor windows for the target events:

• Classes: bike crash · pedestrian slip (with quick recovery → a negative) · fall on stairs · hard-sit / phone-drop (hard negatives) · normal activity.
• Signals: accelerometer + gyroscope (+ barometer for the stairs altitude drop, speed/GPS for bike).
• Domain randomisation (the important part — this is what closes the sim-to-real gap): randomise sensor noise/bias/quantisation, sampling jitter, and phone placement/orientation (pocket / hand / bag), plus body/device variance.
• Emit a common, documented format that the benchmark and training scripts consume.

Acceptance criteria

• Generates each class with tunable parameters and a seed (deterministic given the seed).
• Documented schema; a few example traces committed.
• Lives alongside / extends app/test/sim/.

Guardrails / notes

• Purely synthetic — no real user data. This is the off-device raw-trace stand-in used for training/eval only; on-device the LivenessSource boundary still emits only a verdict.
• Generating plausible data is easy; generating data that transfers to a real phone is the point — lean hard on domain randomisation, and don't over-claim realism (the honest-metrics rule).

Why this is a good first issue / help wanted

Self-contained, no GPU or ML depth needed, and it unlocks the benchmark and training work. Part of the on-device acute-detector epic.

[wrokin]

🤖 wrokin triage

Request for a synthetic IMU data generator that produces labeled fall/crash scenarios with domain randomization to bridge the sim-to-real gap.

Suggested priority: high
Labels applied: enhancement, good first issue, help wanted
Possible duplicate: no

_{Model: deepseek/deepseek-v4-pro · your key, your model (BYOK)}