Controller Learning v0.1.1

Controller Learning v0.1.1 closes the everyday Controller-author loop around the unchanged benchmark 0.1.

Highlights

• New evaluate-controller Pixi command for trusted plugins on Level 0 or ordered Validation prefixes, with deterministic rows/seeds, explicit CPU or MJX-Warp scope, transactional CSV/JSON output, and optional same-rollout trajectory capture.
• New replay Pixi command for strict canonical trajectory loading, deterministic no-overwrite PNG output, and optional interactive public-observation playback without simulation.
• New evidence-derived result interpretation with success/lap-time, speed, and lateral-error views rebuilt deterministically from seven hash-pinned accepted M8 CSV/NPZ artifacts.
• Portfolio-first README, hosted MkDocs site, Controller workflow, stability policy, contribution guide, changelog, and citation metadata.
• Corrected PPO training documentation now includes the required run ID.

Start the Controller loop

pixi install
cp -R controllers/template controllers/my_controller
pixi run sim -- --controller controllers/my_controller --level-id 0 --render
pixi run evaluate-controller -- --controller controllers/my_controller --run-id my-level0 --split level0 --capture-row 0
pixi run replay -- runs/evaluations/my-level0/selected_replays/row_000_trajectory.json --overview runs/evaluations/my-level0/overview.png

• Hosted documentation
• Controller workflow
• Benchmark interpretation
• Stability policy
• Changelog

Verification and benchmark boundary

The release passed 1,108 local CPU tests, all 69 GPU tests on an NVIDIA GeForce RTX 5070 Ti Laptop GPU, strict documentation/actions/package checks, an informal MJX-Warp Validation-only evaluator smoke, and an anonymous fresh-clone Level 0 evaluate/replay workflow. The tagged commit passed GitHub CPU CI and the Node.js 24 Pages deployment.

All 14 frozen PID/MPC/PPO files, all 24 accepted M8 outputs, and five additional protected inputs retain their v0.1.0 SHA-256 identities. Formal M8 Test was not executed. The accepted m8-final-v0-1-002 result and benchmark 0.1 are unchanged.

This release points to source commit d2e1161. Linux x86-64 remains the only tested platform; formal GPU workflows require a compatible NVIDIA GPU.

Released under the MIT License.

Controller Learning v0.1.0

Controller Learning v0.1.0 is the first public release of the GPU-parallel race-car Controller benchmark.

Highlights

• Physical four-wheel MuJoCo vehicle with native MJX-Warp GPU batching.
• Procedurally generated, versioned Level 0/1 Track pools.
• One public Challenge and Controller interface shared by PID, MPC, PPO, and user plugins.
• Pixi-locked Linux/Python 3.11 workflow, strict CPU CI, versioned GPU evidence, and 2D replay.
• English Quick Start, Controller tutorial, API, evaluation, and reproducibility documentation.

Formal benchmark 0.1

Rank	Controller	Test success	Mean successful lap	Compute P99
1	PID	20/20	88.085 s	0.340 ms
2	MPC	20/20	102.563 s	43.902 ms
3	PPO	19/20	23.913 s	0.281 ms

Ranking compares success rate first, then mean successful lap time. All three Controllers used the same fixed 20 Test Tracks, seed identities, four-wheel MJX-Warp environment, public observation/action boundary, and 50 ms soft deadline protocol.

• Canonical M8 report
• Controller comparison CSV
• Canonical Test row-0 comparison
• M7 PPO training curve
• M7 PPO Validation replay

Reproducibility and scope

The accepted M8 evaluation was executed from source commit 6095481. This release points to 65a0fde, which publishes the frozen result and records the public state.

Formal attempt 001 stopped after Test loading but before environment creation, reset, step, Controller construction, or any performance result. Its zero-episode infrastructure failure is retained in the canonical disclosure. The explicitly authorized attempt 002 is the accepted result; no third official attempt is permitted.

Linux x86-64 is the only supported v0.1 platform. Formal GPU workflows require an NVIDIA GPU and use MJX-Warp. macOS, native Windows, and WSL2 are not supported or tested in v0.1.

The canonical distribution is the GitHub source tree with its Pixi lockfile. Start with git clone, pixi install, and the commands in the README. The final local suite passed 1,086 CPU tests and 69 GPU tests; the tagged commit also passed clean-checkout GitHub CPU CI.

Released under the MIT License.