Research Scope and Limits
SidewalkPilot is a solo-built research and learning platform. It is not a product, a road-legal vehicle, or a certified autonomous system. This boundary applies to every model, demonstration, and result in the documentation.
What the Project Is
SidewalkPilot is a physical RC-scale platform used to study the complete camera-steering engineering loop:
- Collecting and correcting real field data;
- Training and comparing custom neural networks;
- Deploying GPU inference to a Jetson Orin Nano: Series 1/2 through PyTorch CUDA and Series 3/4 through ONNX Runtime CUDA;
- Controlling real motors and steering from a Raspberry Pi 5;
- Keeping manual takeover and deterministic LiDAR braking independent of the model;
- Recording telemetry on a separate Zero 2 W dashboard;
- Turning field failures into the next data or code revision.
The value is the integration and iteration, not a claim that one neural-network score establishes general autonomy.
What Has Been Demonstrated
- Physical camera-to-steering runs under operator supervision.
- v3.4 handled every shadow case presented in one July 13 comparison and was selected over v3.4b/v3.3/v3.3b for current use.
- A published 81,237-image real Series 3/4 dataset supports reproducible training work.
- Fifty-two checkpoints can be decoded and compared on one frozen challenge subset.
- LiDAR software can govern throttle and command emergency braking without choosing steering direction.
- Manual steering remained responsive in a hardware retest with Jetson Orin Nano powered off after network work moved outside the control loop.
- Six v4.0 models completed training, ONNX export, offline evaluation, runtime compatibility checks, and supervised field testing.
- Six v4.1 correction models completed training, ONNX export, and offline evaluation.
Evidence Limits
- The July 13 v3.4 result is qualitative. Exact route, weather, takeover count, and clip identifiers were not preserved.
- The v4.0 field result is qualitative rather than a repeated route-controlled benchmark. v4.1 has not yet completed live integration or physical testing.
- Offline metrics do not prove obstacle avoidance, pedestrian response, recovery from a road-edge error, or operation outside the test distribution.
- The LiDAR policy has automated software tests, but the latest configuration still needs a preserved physical test record.
- GPS/navigation code exists, but it does not make the car safe for unattended route execution.
- The project has not received third-party safety validation or certification.
Operating Boundary
- A human operator remains present with the controller and immediate stop authority.
- Tests are bounded, directly supervised, and do not authorize autonomous public-road operation.
- The car is not operated unattended.
- Autonomy is not armed when required sensors, manual controls, model inference, or braking checks fail.
- LiDAR does not steer around obstacles. It slows, holds, or emergency-brakes in the center corridor.
- New models remain experimental until a documented field comparison promotes them.
The platform is not authorized for public-road autonomy, unattended use, uncontrolled pedestrian operation, all-weather driving, or safety-critical service. A selected sidewalk test route does not become a closed course merely because it is familiar. Testing stops when people, traffic, weather, visibility, hardware, or system freshness violate the planned setup.
Reporting Rules
- Label a result as offline, bench, or field evidence.
- Distinguish exact measurements from qualitative operator observations.
- Do not convert “worked in the tested cases” into “works in all cases.”
- Do not call a trainer-selected checkpoint the best driving model without field evidence.
- Keep planned work separate from implemented work.
- Preserve the model hash, route, conditions, clips, logs, and takeovers for future field claims.
See Safety Overview, Operating Limits, and Evidence Map.