Overview
This project explores autonomous racing-line control through a physics-based simulator and an imitation-learning pipeline. The work combines behavior cloning, corrective data collection, and failure-focused evaluation rather than treating aggregate performance as the only measure of success.
Problem
A policy can perform well on familiar track segments while failing sharply under unseen geometry, high-curvature transitions, or unstable vehicle states. Building a useful controller therefore requires both learning a racing behavior and understanding where that behavior breaks.
Dataset / Inputs
- Expert driving trajectories generated in simulation
- DAgger-style correction data from observed policy failures
- Vehicle state features including velocity, yaw rate, and slip angle
- Look-ahead geometry and curvature observations
- Familiar and unseen track segments for robustness evaluation
System Architecture
- Physics-based racing environment
- Trajectory and correction-data generation workflow
- Behavior-cloning training pipeline
- DAgger-style iterative correction loop
- Evaluation harness for recurring failure segments and zero-shot transfer
Methodology
The initial policy learned from expert demonstrations through behavior cloning. Repeated failure segments were isolated during evaluation, then used to collect targeted correction data. Observation design was revised to include look-ahead geometry, curvature, yaw rate, velocity, and slip angle so the policy received more useful context before difficult transitions.
What I Built
- Physics-based racing simulator
- Imitation-learning pipeline for behavior cloning
- DAgger-style correction-data workflow
- Robustness and zero-shot transfer tests
- Failure-segment analysis and observation redesign
Results / Proof Points
- Identified recurring failure segments instead of relying only on aggregate lap behavior
- Tested transfer to unseen track geometry
- Redesigned the observation space around vehicle dynamics and upcoming curvature
- Built a repeatable correction loop for improving weak policy behavior
Business Value
The project demonstrates a practical model-development workflow: build a system, test it outside its easiest conditions, isolate failure patterns, and redesign data or observations around the evidence.
Limitations
The current project is simulation-based and remains in active development. Public performance metrics and repository evidence will be added after the evaluation protocol and report are finalized.
Next Steps
- Formalize lap-time, stability, and intervention metrics
- Compare behavior cloning and correction strategies under the same evaluation suite
- Expand testing across additional track geometries
- Publish the project report, visualizations, and implementation evidence