Image Processing Project

Feature-Based Image Processing and Reinforcement Learning for CarRacing-v3. Compact road-geometry features are evaluated with PPO and SAC.

Demo Videos

PPO Completed Baseline

938.87 +/- 7.86 @ 500,000

Completed 500K baseline.

SAC Fast-Result Branch

938.51 +/- 4.88 @ 400,000

Partial 400K best-checkpoint, not a completed 500K run.

PPO Completed Baseline

Validated eval: 938.87 +/- 7.86 @ 500,000

Completed 500K baseline preview from the saved best demo clip.

SAC Fast-Result Branch

Validated eval: 938.51 +/- 4.88 @ 400,000

Partial 400K best-checkpoint preview, not a full-length SAC run.

Method at a glance

Feature-based visual-control pipeline: frame, perception, geometry, state vector, temporal context, and control

RGB frame from Gymnasium CarRacing-v3.
HSV road mask and ray/radar geometry features.
16D base observation with a four-frame stack for a 64D policy input.
PPO and SAC MLP policies used as downstream evaluation methods.

Interpretation

PPO is the completed 500K baseline. SAC is a partial 400K fast-result branch and is not presented as outperforming PPO. The comparison is not compute-equivalent, this project package does not include raw-pixel CNN policy training, and it makes no environment-completion claim for CarRacing-v3.