This release documents and publishes the hardened pipeline after the large strategy update.
- Added route-aware terminations:
offroute,noprogress, androute_complete. - Added strange-situation guards for stale telemetry, paused/menu states, frozen capture frames, teleport jumps, long GPU-stall ticks, and slow-but-advancing motion.
- Reworked AutoDrive recovery around the observed FH6 behavior:
- far off-road can show a Fast Travel Warning / transfer prompt
- on-road stuck states usually have no prompt and AutoDrive drives back
- confirm
Ais only sent while the car is positionally frozen, which fixes the case where coast-down after a crash blocked the confirm press - a drivable parallel road is accepted instead of teleporting a fine car
- when every recovery option is exhausted the run pauses for manual help instead of crashing the training process
- Added
forza_fullDreamer config for larger world-model training when GPU headroom is available. - Added
train_dreamer.py --configsoforzaandforza_fullare selectable. - Redrew the README diagrams (hero, architecture, recovery loop, hardening stack) with a consistent visual language.
- Added stress tests for NaN/inf handling, malformed packets, teleport jumps, route seams, braking-vs-impact, route ends, offroute, noprogress, and stuck edge cases.
- Updated the public README and SVG diagrams to match the current stack.
python -m unittest discover -s tests -v
python -m py_compile recovery.py recovery_demo.py forza_rl_env.py train_dreamer.py reset_test.py offline_pretrain_dreamer.py
At publication time, the suite contains 55 tests.
forza_fullshould only be used after reducing FH6 GPU pressure and verifying that VRAM does not spill into shared system memory.- A waypoint must be pinned for AutoDrive recovery.
- Teleport recoveries keep training alive but are not saved as dynamics demos.
Initial public-ready snapshot.
- Real-time Forza Horizon driving environment built from screen capture, Data Out telemetry, and virtual Xbox controller input.
- DreamerV3 integration through a reproducible vendor patch for
NM512/dreamerv3-torch. - Low-VRAM workflow with offline replay pretraining.
- Centerline-aware reward for route progress.
- Crash, stuck, flipped and off-road detection.
- AutoDrive recovery demonstrations for smooth non-teleport recoveries.
- Public documentation and SVG diagrams.