Project Contributions and Scope

This page separates project-specific work from established machine-learning and robotics patterns. It does not claim that a familiar technique becomes a new invention because it was implemented on SidewalkPilot.

The direct historical reference is NVIDIA's 2016 End to End Learning for Self-Driving Cars, which trains a convolutional network from road images paired with human steering commands. SidewalkPilot adopts that supervised camera-to-steering framing but does not reproduce the paper's vehicle, routes, dataset, compute, or validation protocol. Its results are therefore not compared numerically with PilotNet.

Commercial sidewalk robots provide problem-setting context, not a benchmark. This repository does not contain controlled sensor, cost, speed, or safety comparisons against commercial platforms and makes no claim of matching them.

1. Failure-Driven Sidewalk Dataset and Test Loop

Status: implemented, with bounded field evidence. The project collects its own camera/control data, identifies failure conditions, adds targeted examples, retrains, and returns to the car for comparison. The July 13 test recorded v3.4 as the best of the four Series 3 candidates tested in normal turns and the shadow cases presented that day.

This is a project contribution because the dataset, hardware integration, failure clips, and iteration record were produced for this car. It is not evidence that v3.4 is universally shadow-robust; the route identifier, quantitative takeover count, and repeated-trial record for that comparison were not preserved.

2. Hybrid Steering Head Applied to Sidewalk Driving

Status: implemented in v3.1 through v3.4b. SidewalkPilotV3 predicts nine steering-class logits, nine local offsets, and throttle. Combining a discrete choice with a continuous offset is an established modeling pattern. The project-specific design choice is applying that pattern to a steering dataset where straight-heavy labels can hide weak turn recall.

The architecture is therefore described as an application and engineering experiment, not as a new class-plus-offset algorithm.

3. Series 4 Temporal Experiments

Status: v4.0 field-tested; v4.1 evaluated offline. Series 4 compares causal history, future supervision, and combined past/current/future targets while keeping the image backbone and 18-value steering head comparable. Future targets are training supervision, not unavailable future inputs at deployment.

The v4.0 comparison demonstrated why closed-loop testing matters: PC/PCF ranked strongly offline but echoed earlier predictions on the car. Image-only v4.0f was viable and mixed against v3.4. The six v4.1 models test corrections to that failure and remain pending live integration and field testing.

4. Integrated Jetson Orin Nano, Raspberry Pi 5, Sensors, and Dashboard

Status: implemented engineering. The Jetson Orin Nano runs every steering-model family on its GPU: Series 1/2 through PyTorch CUDA and Series 3/4 through ONNX Runtime CUDA. The Raspberry Pi 5 owns controller input, sensors, steering and motor output, logging, and dashboard telemetry. The Zero 2 W renders the USB-linked dashboard. Splitting hardware I/O from heavier inference is an established systems pattern; the contribution is the working integration and debugging record on this vehicle.

5. Evaluation Beyond MAE

Status: implemented evaluation practice. The report reads MAE alongside Bal9, turn exact, turn +/-1, straight exact, signed error, and confusion matrices. Macro recall and class-aware evaluation are standard methods. Their value here is practical: they expose straight collapse that one aggregate error number can conceal.

Evidence Standard

  • Implemented means the code, model, or other stated evidence exists.
  • Offline result means it was measured on the stated frozen evaluation set.
  • Field observation means it occurred in a bounded physical test.
  • Planned means it is not complete and is not presented as a result.