Feature-Based Image Processing and Reinforcement Learning for CarRacing-v3
Joe / Student ID 1103820
An inspectable image-processing pipeline converts CarRacing-v3 RGB frames into compact road-geometry features for PPO/SAC control evaluation. Instead of training a raw-pixel CNN policy, the package emphasizes an explicit, inspectable, and lightweight visual state representation.
PPO and SAC demo videos
Open the demo page to view pauseable PPO and SAC preview clips with conservative result notes.
Open PPO + SAC demo page
·
Full video gallery
| Branch | Status | Validated eval | Best parsed eval |
|---|---|---|---|
| PPO completed baseline | Completed 500K baseline | 938.87 +/- 7.86 @ 500,000 | 939.53 +/- 4.09 @ 480,000 |
| SAC fast-result branch | Partial 400K best-checkpoint / fast-result branch | 938.51 +/- 4.88 @ 400,000 | 938.51 +/- 4.88 @ 400,000 |
Conservative interpretation. PPO is a fully completed 500K baseline. SAC is a strong but partial fast-result branch whose validated number is the 400K checkpoint. The PPO/SAC comparison is not compute-equivalent, and SAC is not presented as outperforming PPO. The project makes no environment-completion claim.
Metric context. In CarRacing-v3, reward is tied to track-tile coverage with a per-frame time penalty, so scores in the 930s should be read as high-return evaluation episodes under the saved protocol. This package still does not make a general environment-completion or compute-equivalent PPO/SAC claim.
- RGB frame from Gymnasium CarRacing-v3.
- HSV road mask and compact road-geometry extraction.
- Ray/radar features summarized into a 16D base observation.
- Four-frame temporal stack produces a 64D policy input.
- Standard Stable-Baselines3 PPO/SAC MLP policies evaluate the representation.
This project package does not include raw-pixel CNN policy training.
| Material | Link |
|---|---|
| Final IEEE-format technical report PDF | 1103820_Joe_Image_Processing_Project_V2_IEEE-format report.pdf |
| Overleaf / LaTeX source package | final/overleaf/Image_Processing_Project_Overleaf_Package.zip |
| Final presentation (PDF) | final/slides/Image_Processing_Project_Final_Presentation.pdf |
| Final presentation (PPTX) | final/slides/Image_Processing_Project_Final_Presentation.pptx |
| PPO notebook | final/notebooks/Final_PPO_Baseline_CarRacing_v3.ipynb |
| SAC notebook | final/notebooks/Final_SAC_Fast_Result_CarRacing_v3.ipynb |
| Result summary | final/docs/RESULT_SUMMARY.md |
| Figures | final/figures/ |
| Video gallery | final/videos/README.md |
final/ public-facing final package
report/ final IEEE-format report PDF
overleaf/ LaTeX source package for the report
slides/ final presentation (PPTX / PDF)
notebooks/ PPO and SAC notebooks with saved outputs
figures/ report and slide figures + demo poster
tables/ CSV result evidence
videos/ demo MP4s, previews, manifest, gallery README
logs/ training and evaluation logs
docs/ result summary, run instructions, dataset notes
validation/ static validation report
archive/ preserved history, not the current project package
Start with final/README.md, the
result summary, and the
Final IEEE-format technical report PDF. The
notebooks are kept in safe report/evaluation mode and include saved output
evidence; they do not require retraining to inspect. Do not rerun training
unless you intentionally change the project scope.
For local/static inspection guidance and Colab-safe report/evaluation mode,
see final/docs/RUN_INSTRUCTIONS.md.
Key evidence paths:
- Logs - headline metrics are parsed from these.
- Tables - checkpoint and summary CSV evidence.
- Validation report - static package checks.
- Run instructions - reading/inspection guidance.
- The SAC fast-result branch is a partial 400K checkpoint, not a full-length 500K run.
- SAC is not presented as outperforming PPO.
- PPO and SAC are not a compute-equivalent comparison.
- This project package does not include raw-pixel CNN policy training.
- The project makes no environment-completion claim for CarRacing-v3.
- HSV-style visual preprocessing can be sensitive to rendering and track appearance changes.
The original V1 PPO-only submission and the dated V2 build snapshot are preserved
under archive/ for traceability. They are not the current public
project package. V1 results must not be confused with the V2 PPO/SAC results
above.