Overview
This project studies autonomous driving in a physics-based racing simulator. It combines expert demonstrations, behavior cloning, corrective data, and matched evaluations to understand where a policy succeeds and where it fails.
The operational Monza driver remains the frozen Epoch 3 navigation baseline. Later pedal-control experiments are evaluated candidates, not replacements for that baseline.
Problem
Completing a familiar lap is different from driving reliably under changed starting conditions, and reliable navigation is different from improving lap time. The evaluation keeps those questions separate: valid finish crossing and ordered checkpoints first, safety and regressions next, speed after that.
What I Built
- A Python simulator with world-space Monza geometry, vehicle dynamics, telemetry, checkpoint gates, and fixed physics steps.
- Expert demonstrations, behavior-cloning training, and failure-focused corrective data collection.
- Observations combining vehicle state with look-ahead geometry, curvature, yaw rate, velocity, and slip angle.
- A hash-verified baseline release and an interactive viewer for manual, teacher, and learned drivers.
- Matched-start evaluations and a read-only evidence checker that rejects incomplete, unsafe, duplicated, or inconsistently paired results.
- A bounded pedal-residual experiment that preserves the baseline's steering and tests oval regressions alongside Monza behavior.
Evaluated Results
The October 21, 2026 recorded evaluation of the Epoch 10 imitation initialization completed 60 of 60 cases safely: 20 Monza fixed-start cases, 20 randomized-start cases, and 20 oval cases.
- Mean Monza gain over the frozen BC reference: about 9.753 seconds.
- Mean Monza gain over the hand-set fixed correction: about −0.012 seconds; the imitation candidate was slightly slower.
- Oval regression: unchanged in the evaluated cases.
- Training for this imitation candidate: 520 supervised updates; zero PPO updates.
- Promotion: not promoted; Epoch 3 remains the operational baseline.
The fixed correction is a hand-set controller. The candidate learned constant residual targets and still depends on a predefined Monza distance gate. This result is not evidence of adaptive racing skill, reinforcement-learning improvement, or superiority over that fixed controller.
The simulator's separate frozen navigation benchmark records a teacher lap near 298.57 seconds and an Epoch 3 mean near 325.53 seconds. Neither controller nor its recorded racing line is asserted to be globally optimal.
Evaluation Limits
The 60 imitation cases were fresh candidate runs compared with earlier hash-preserved, configuration-matched BC and fixed-controller results; those controls were not rerun concurrently. Fixed-start seeds share a deterministic spawn, while randomized starts provide separate perturbations.
These results apply to the frozen simulation and configuration. They do not establish real-vehicle safety, general track transfer, or an optimized racing policy. Later experiment-preparation checks are infrastructure evidence, not additional driving results.
What I Learned
A learned controller can reproduce a useful correction without discovering a better policy. Reporting the fixed-controller comparison prevents a gain over the original baseline from being mistaken for a new learning breakthrough.
Next Steps
Preserve the operational baseline and completed evaluations. Any further optimization needs a separately reviewed experiment and its execution gates satisfied before training or evaluation; eligibility for review is not promotion.