Auto Thinker Engine v1

Auto Thinker Engine v1

Version 1.0.0 is the first tagged source and research-documentation release of
Auto Thinker Engine. It captures the existing CARLA driving research and the
paused camera-navigation work, together with a reproducible CPU software demo.

Included

• JEPA video encoders, DreamerV3 integration, and the CNN/JEPA/V-JEPA 2 comparison.
• Multi-camera navigation, perception, data tooling, scoring and planning studies.
• Frozen experiment configurations, software tests, tracked vendor patches,
  corrected research reports and the conference paper.
• README About section, animated execution demo, captured transcript and regeneration script.
• Package metadata and lockfile aligned at 1.0.0; release tag v1.

Scope and limits

This is a research snapshot. The project remains paused at its 23 October 2026
checkpoint. Camera safety, rule understanding and predictive-planning gates remain
open; the registered final Town05 driving evaluation has not run. The release does
not declare a qualified driver or completed M0–M6 research goal.

The README GIF records software checks on synthetic observations. It does not show
CARLA footage or establish driving quality. Datasets, recordings, model weights,
simulator binaries and navigation runtime artifacts under outputs/ are excluded.
Clone recursively and follow the README/runbooks to install the required runtime.
GitHub's generated source archives do not contain the contents of Git submodules.

Third-party code and model weights retain their upstream licenses and usage terms.

Verification

The GIF generator requires successful execution of the runtime doctor and all
eight existing synthetic integration tests, including CNN/JEPA optimizer checks
and checkpoint save/load. Its transcript is checked in beside the GIF.

The historical full-suite result in the README (559 passed, 3 skipped) belongs to
the original research machine. It is not a claim that a fresh source-only checkout
can run every test without the optional navigation vendor, data and dependencies.
The release preparation checks and current environment limits are recorded in
the demo provenance file.

On the release-preparation CPU checkout, the full suite returned 534 passed,
18 skipped and 10 failed, due to unavailable ignored navigation artifacts,
the optional navigation vendor and numba. The eight demo integration tests,
script lint/compilation and offline lockfile check passed. The local Python 3.10 /
PyTorch 2.4.1 CPU environment differs from the pinned GPU research environment.