Field Evaluation Overview
Field evaluation checks behavior that an image-label report cannot measure: Raspberry Pi 5–Jetson Orin Nano latency, steering smoothness, shadow response, mechanical drift, operator takeovers, and LiDAR braking on the assembled vehicle.
Offline Versus Field Evidence
The common evaluator writes docs/steering_eval_current_labels.json and
docs/steering_model_report.pdf for 52 checkpoints on a frozen 6,952-frame
Series 3/4 subset. Bal9 and turn metrics help reject center-collapsed candidates;
MAE, median error, and signed error add magnitude and bias context.
Those values do not select a field baseline by themselves. A model must load through its matching live Jetson Orin Nano path, preserve manual response, and complete a supervised field comparison under the condition it is intended to improve.
Current Field Record
The July 13 comparison is an operator-observed, bounded field result:
- v3.4 handled every shadow case presented in that run and is the current field-selected baseline;
- v3.4b was slightly worse;
- v3.3 was worse than v3.2;
- v3.3b was much worse than v3.2b.
This record does not contain enough repeated-route quantitative measurements to claim a universal success rate. In a later supervised v4.0 comparison, v4.0f was viable and complementary with v3.4, v4.0g was worse, and the PC/PCF models echoed prior steering predictions. The v4.1 correction models remain offline-only.
Run Record
A reproducible field comparison should retain:
- Model version and model file hash;
- Route, surface, lighting, weather, battery, and payload;
- Start/end time and distance;
- Takeover count and cause;
- Runtime CSV (nominal 10 Hz, 46 columns);
- Video/clip identifiers;
- AEB state and any LiDAR intervention;
- Pass, warning, or failure decision made before examining the next model.
The operator keeps the Xbox controller ready. Steering, gas, or brake input
cancels autonomy through cancel_autonomous_mode().
What the Run Must Observe
- Manual takeovers: count each autonomy cancellation and record its cause, location/node, previous autonomous duration, and whether it prevented a curb, grass, road, obstacle, or control-quality failure.
- Steering smoothness: distinguish stable corrections from oscillation, delayed response, or abrupt class-boundary movement.
- Curb and road risk: record approach direction and whether recovery occurred before the operator intervened.
- Shadow behavior: test ordinary turns and turns under distinct shadow geometry; do not generalize one successful route to every lighting condition.
- Evening/night behavior: remains a separate, higher-risk condition and is not approved by daylight results.
- LiDAR: record governor, hold, and emergency actions separately from model errors. LiDAR does not steer.
The historical 0.5-mile/five-intervention run is journey context, not a formal current-model benchmark because its complete date, route, and run record were not preserved.