Echo 1 — Safety engine

The safety engine is the heart of Echo 1. It is pure Python with no dependency on the language model, the cloud, or the internet — so it keeps working even if everything else fails.

Obstacle detection

YOLOv8n on the Hailo-10H runs at 30 FPS (~17 ms inference) and names hazards from the COCO set. Proximity is estimated from the bounding-box area fraction; direction is mapped to a clock position across the field of view. Only one alert is emitted per direction — the most severe wins.

Drop-off & step detection

The down-facing LiDAR auto-calibrates a floor baseline. A reading much farther than baseline means a drop-off (“Stop. Drop-off ahead.”); much closer means a step up (“Caution, step up ahead.”). A confirm-streak filter and slow drift tracking suppress false positives.

Alerting

The Alert Manager applies cooldowns and routes warnings through the Audio Manager at safety priority, interrupting any lower-priority speech. During recording, only danger-level safety speech plays so the mic doesn't capture the speaker.