drivingbench-traces / README.md
aditya-ramabadran's picture
DrivingBench traces
15b8369
|
Raw History Blame Contribute Delete
3.86 kB
metadata
license: cc-by-4.0
pretty_name: DrivingBench traces
tags:
  - robotics
  - autonomous-driving
  - agents
  - benchmark
size_categories:
  - n<1K

DrivingBench release

Language models driving a real car. Each attempt below is one model's try at a mini-cone course in an empty parking lot, at walking pace, with a human in the driver's seat ready to brake. The model saw the road through two cameras and drove with three tools (observe, set_motion, stop_now); every command, the reason the model gave for it, the car's telemetry and the video are here, aligned to the frame. The harness that recorded this is open source: https://github.com/aditya-ramabadran/drivingbench_harness_v1.

Created by: Simon Mahns, Tobias Gessler, Aditya Ramabadran.

Layout of the September trial: <collection>/<model group>/<attempt>/ with road.mp4, wide.mp4, road-annotated.mp4, calls.srt, timeline.json, frames/, README.md; each group has the trial's transcript.md. The v1.1 evaluation is laid out as <model>/<n>/ (its README). manifest.json indexes everything.

v1.1-evaluation/: DrivingBench v1.1

Two cone courses on one lot held out from all training: Course A (≈ 141 m) and Course B (≈ 92 m), on a Tesla Model 3 with a comma. Chat models ran three independent trials of up to three same-context attempts per course, action models five independent episodes per course, and a human one trial of three attempts per course. Each attempt has its road and wide video, timeline of commands and telemetry, GPS track, transcript, the frames a chat model saw, and token usage. Scores are in results.json and the course geometry in course-a.json and course-b.json; v1.1-evaluation/README.md gives the scoring rule, the layout and how to re-score every attempt (scoring/score.py). We went through every v1.1 video by hand: there were no people or licence plates to blur. Results: https://drivingbench.com.

  • flux-3-action/: FLUX 3 Action (10 attempts)
  • pi0-7/: π0.7 (10 attempts)
  • pi0-5/: π0.5 (10 attempts)
  • dream-1/: Dream-1 (10 attempts)
  • gr00t-n1-7/: GR00T N1.7 (10 attempts)
  • xiaomi-robotics-1/: Xiaomi Robotics-1 (10 attempts)
  • gpt-6-astra/: GPT-6 Astra (18 attempts)
  • gpt-6-1-sol/: GPT-6.1 Sol (18 attempts)
  • gpt-5-6-sol/: GPT-5.6 Sol (18 attempts)
  • claude-opus-5-5/: Claude Opus 5.5 (18 attempts)
  • claude-fable-5-2/: Claude Fable 5.2 (18 attempts)
  • claude-fable-5-1/: Claude Fable 5.1 (18 attempts)
  • gemini-4-carbon/: Gemini 4 Carbon (15 attempts)
  • grok-4-6/: Grok 4.6 (18 attempts)
  • human-baseline/: Human baseline (6 attempts)

2026-09-17-cone-course-trial/ — Cone course trial, Sep 17 2026

One trial per model, up to three in-context attempts each, medium reasoning. Course: a backwards-U parking lot marked with lime/orange mini-cones and tall orange cones, finishing in a blue mini-cone parking area. An attempt ends when the supervising driver brakes for a mistake or an imminent collision; the operator then sends a fixed reflection prompt and, for the next attempt, a fixed continuation prompt in the same chat. GPT-6 Astra completed the course on its second attempt; no other attempt finished.

  • gpt-6-astra/ — GPT-6 Astra (Codex, medium reasoning): F S · attempt-1--fail--24137e, attempt-2--success--5c4479
  • gpt-5.6-sol/ — GPT-5.6 Sol (Codex, medium reasoning): F F F · attempt-1--fail--5ceee7, attempt-2--fail--627be4, attempt-3--fail--48ff3a
  • claude-fable-5.1/ — Claude Fable 5.1 (Claude Code, medium reasoning): F F F · attempt-1--fail--18ade2, attempt-2--fail--926568, attempt-3--fail--ca2652
  • grok-4.6/ — Grok 4.6 (Cursor, medium reasoning): F F F · attempt-1--fail--f1762a, attempt-2--fail--503d24, attempt-3--fail--667691