언어 모델이 주차장에서 실제 차를 운전하는 작은 마구 3개의 공구를 통해서 속도, 인간은 발을 가진 운전사의 좌석에 앉습니다 브레이크에. 그것은 comma 장치에서 실행 지원된 도요타에 openpilot, 및 MCP (Codex, Claude Code, Cursor)를 말하는 모든 채팅 앱에 차를 노출 하나의 서버, drivingbench_sandbox, 세 가지 도구: observe, set_motion, stop_now. 모든 모델은 말한다 고, 그리고 모든 차는, 이다 당신이 재생할 수있는 추적으로 기록.

이것은 닫히는 과정을 위한 연구 소프트웨어입니다. 이전의 다음 섹션을 읽으십시오 다른 것.

부인과 안전

**당신의 위험에 사용. ** 이 소프트웨어는 스티어링, 가속 및 명령 언어 모형의 산출에 근거를 둔 모터 차량의 제동. * 이름 모형은 실수를 만듭니다; 이렇게 이 부호, 장치, 차 및 사람 감독. 저자는이 소프트웨어를 제공, 어떤 보증 없이 어떤 손상, 상해, 손실, 벌금, 또는 다른 사람을 위한 책임 없음 그 사용 또는 그것에서 파생 된 모든 것에서 발생하는 결과. 이용안내 그것은, 당신은 운전사와 통신수이고, 당신은 차량을 책임집니다 모든 순간.

DrivingBench Harness

A small harness that lets a language model drive a real car at parking-lot speed through three tools, while a human sits in the driver’s seat with a foot over the brake. It runs on a comma device with openpilot in a supported Toyota, and exposes the car to any chat app that speaks MCP (Codex, Claude Code, Cursor) as one server, drivingbench_sandbox, with three tools: observe, set_motion, stop_now. Everything the model says and does, and everything the car did, is recorded as a trace you can replay.

This is research software for closed courses. Read the next section before anything else.

Disclaimer and safety

Use at your own risk. This software commands the steering, acceleration and braking of a motor vehicle based on the output of a language model. Language models make mistakes; so does this code, the device, the car, and the person supervising. The authors provide this software as is, without warranty of any kind, and accept no liability for any damage, injury, loss, fine, or other consequence arising from its use or from anything derived from it. If you use it, you are the driver and the operator, and you are responsible for the vehicle at every moment.

The authors are not affiliated with, endorsed by, or supported by comma.ai, openpilot, Toyota, or the makers of any model or chat application named here. openpilot’s own terms and safety model apply to the device; this project modifies openpilot in ways its authors never intended or reviewed.

Do not trust the software limits. The harness has a speed ceiling, an emergency speed cut-off, command expiry, and cancels motion on brake, pedal, disengagement, or invalid sensors. These are ordinary code paths on an unreviewed system, not safety functions; any of them can fail. The safety measures are the human ones:

• A licensed, sober, attentive driver in the driver’s seat, foot over the brake, hands ready, at all times. The brake or the steering wheel always wins: pressing either cancels the model’s command and disengages openpilot.
• Private property you have permission to use, empty of people and other vehicles, with soft obstacles (cones) only. Never on public roads, never near pedestrians, never with passengers, never at night or in rain until you have far more experience with the system than this README can give you.
• Walking pace. The defaults cap commands at 3 m/s (about 7 mph); keep them there.
• One operator supervises the car and the UI; treat every command in the Drive-tab history as something you personally authorized.
• Do a supervised check of every release before letting a model drive with it (docs/deployment.md), and recalibrate the steering scale on your car (docs/steering-calibration.md).
• Know your local law. Operating a modified driver-assistance system, even on