Safety Overview
SidewalkPilot moves a physical RC-scale vehicle. Its current safeguards reduce specific risks during supervised tests; they do not constitute certification or a complete functional-safety system.
Implemented Layers
- Operator control: while the Xbox controller is connected and the Raspberry Pi 5 loop is responsive, steering, gas, or brake input cancels autonomy. The Share button requests an orderly shutdown. The operator also needs an independent way to cut power.
- LiDAR longitudinal intervention: when AEB is enabled and fresh center- corridor returns are available, the policy can cap forward throttle and request a hard brake at 1.05 m. LiDAR never commands steering.
- Inference freshness: unavailable or stale model results cause the
autonomous path to request a hard stop. The current neural path assigns
confidence
1.0to accepted fresh results, so the confidence field is not a calibrated detector for wrong scenes. - Operating procedure: tests require line-of-sight supervision, bounded
routes, dry conditions, and no autonomous public-road operation. The declared
MAX_AUTONOMOUS_SPEED_MPHis not wired into a measured-speed governor and must not be described as an enforced cap.
Known Gaps
- Stale or empty LiDAR data removes obstacle intervention rather than forcing a stop.
- Software override depends on a connected controller and responsive process.
- Configured LiDAR thresholds do not prove physical stopping distance.
- No arbitrary-pedestrian, all-weather, or unattended-operation claim is made.
- Quantitative false-trigger, disconnect, stopping-distance, and override-latency records still need controlled physical tests.
Evidence Standard
Code and unit tests establish the configured arbitration logic. Physical claims require a preserved setup, payload, speed, route, logs, video, and pass/fail record. The July 13 model comparison selected v3.4 in the cases presented, but it was not a safety certification or complete route benchmark.