Skip to content

rolling_buffer_spin_detection.md

Latest commit

 

History

History
78 lines (58 loc) · 2.74 KB

rolling_buffer_spin_detection.md

File metadata and controls

78 lines (58 loc) · 2.74 KB
Raw

Rolling Buffer Capture and Spin Detection

OpenFlight uses the OPS243-A rolling buffer for production shot capture. Spin estimation from that capture remains experimental.

Capture pipeline

  1. The SEN-14262 detects impact and triggers the OPS243-A through HOST_INT.
  2. The radar freezes and returns 4,096 I samples and 4,096 Q samples.
  3. RollingBufferProcessor extracts ball and club speed from short FFT windows.
  4. An overlapping timeline separates the club, impact, and ball regions.
  5. An ungated multitaper estimator records an experimental spin candidate.

The normal kiosk command uses this pipeline:

scripts/start-kiosk.sh

One-time radar setup

The OPS243-A firmware changes the HOST_INT behavior when rolling-buffer mode is entered at runtime. Save the mode to flash once, then power-cycle the radar:

uv run python scripts/hardware-test/test_rolling_buffer_persist.py --setup
# Disconnect power from the radar for at least three seconds, then reconnect it.
uv run python scripts/hardware-test/test_rolling_buffer_persist.py --test

The runtime then starts in the persisted GC rolling-buffer mode without re-entering it. See Sound Trigger Wiring for the recommended direct hardware trigger.

Current defaults

Setting Value
OPS243-A mode Persisted rolling buffer (GC)
Sample rate 30,000 samples/s
Capture 4,096 I + 4,096 Q samples
Capture duration About 136.5 ms
Pre-trigger split 16 of 32 blocks
Speed window 128 samples
FFT size 4,096
Overlap step 32 samples (937.5 windows/s)
Production trigger SEN-14262 GATE to OPS243-A HOST_INT

Use --buffer-split post-heavy when more post-impact data is needed, or --buffer-split pre-heavy when the club approach matters more:

scripts/start-kiosk.sh --buffer-split post-heavy

Experimental spin output

Every valid capture runs the ungated multitaper estimator. Its candidate:

  • is labeled EXPERIMENTAL in the UI;
  • is logged as spin_method: "multitaper_ungated";
  • is constrained to 1,500–11,000 RPM; and
  • does not drive spin-adjusted carry.

Treat individual values as evaluation data, not ground truth. Multipath and capture geometry can produce plausible but incorrect candidates, and current blind validation is not accurate enough for production carry calculations.

For offline comparison and estimator work, use Spin Replay and Diagnostics. Raw captures are stored in the session JSONL logs described in the observability guide.

Related guides