DrivingBench Harness v1.1

The harness behind DrivingBench v1.1. The three MCP tools are unchanged; what's around them grew.

New

• Tesla Model 3 (HW3) alongside the Toyota, on openpilot 59fcd340, the source of the comma four release-mici-staging build that supports Tesla's current steering firmware (docs/tesla.md).
• The runner loop (docs/runner.md): a chat model on OpenRouter or Google's Gemini API drives with exactly the public tools and no chat app, or supervises a fast driver: a decision model, or any trained policy (VLA) you serve over a small HTTP contract. Policies send timed curvature and speed plans anchored to their frame's capture, executed by the same controller and limits.
• The simulator (docs/sim.md): the v1.1 course as a Gaussian splat, driven by the real native stack and openpilot's Tesla controllers over a car fitted from its logs, with course progress, cone contacts and the finish scored as on the car. Every client drives it unchanged. Includes the pipeline that builds it from a day of drives.
• A retuned speed loop, cameras encoded only while someone reads them, and per-session uploads of labeled sessions' bulk artifacts to your own Hugging Face dataset.

Upgrading from v1

• openpilot is now the vendor/openpilot submodule: GIT_LFS_SKIP_SMUDGE=1 git submodule update --init --depth 1 vendor/openpilot, then the same for its opendbc_repo (see the README).
• The comma needs the matching openpilot build and a new bundle (docs/deployment.md); re-run drivingbench install on each laptop. The producer refuses v1's saved settings: delete /data/drivingbench-v01/state/settings.json once so the bundle's initial settings seed it again.

Not included: clients for providers' private model APIs (trained policies plug in through the HTTP contract instead), and the simulator's trained splat, which is not distributed yet: build it with the pipeline or run the simulator with --placeholder.

DrivingBench Harness v1

The harness behind DrivingBench v1: AI models drive a comma-equipped Toyota at parking-lot speed through three MCP tools (observe, set_motion, stop_now), with a supervising driver at the wheel.

• One MCP server, drivingbench_sandbox, for Codex, Claude Code and Cursor.
• A native 100 Hz controller inside openpilot v0.11.2 (two modified files) that turns timed steering-percent commands into openpilot's own curvature and longitudinal control, under openpilot's and the car's limits.
• A laptop gateway with Setup, Drive and trace-viewer pages, and every tool call, image and 100 ms of telemetry recorded as replayable traces.

Read the README's disclaimer before using it: this code moves a car.