Oil · Cafeterrace¶
Agent runs
Each mode of this task runs once per run. Est. cost is tokens at list price (data/prices.yml), never a bill; see the RoboPaint runs for every task of a run.
Codex + GPT-6 Luna, reasoning xhigh (ChatGPT login)default run
| Trial | Started | Agent time | Model requests | Tokens in / out | Est. cost | Grade |
|---|---|---|---|---|---|---|
| U✗ 2h 00m 44% | 09-28 13:02 | 2h 00m | 274 | 41.6M / 166k | $0.570 | deterministic 1, grader_error 0, instance_exact 1, instance_ok 1, missing_trajectory 0, n_actions 31570, video_rendered 1 |
| L✗ 1h 41m 44% | 09-29 17:51 | 1h 41m | 97 | 9.5M / 120k | $0.175 | deterministic 1, grader_error 0, instance_exact 1, n_actions 35613, replay_success 0, video_rendered 1 |
Task instruction (upstream)
Paint the picture shown in /app/target.png (a square crop of Vincent van Gogh's Cafe Terrace at Night, 1888) onto the main sheet with brush and oil paints, in at most 441 strokes.
Playback speed
Recorded by us: our reference solution, replayed by the verifier (canvas camera, scene camera, side camera, exact canvas).
What RoboPaint states about this task
| Success criteria | 1. scored at a 1 mm scale against the picture, all after the best single shift of the sheet, at most 5 mm along each axis, inside the central area 8 mm in from the sheet's edges 2. mean CIELAB ΔE <= 10.9 3. p90 ΔE <= 20.0 4. SSIM of L* >= 0.5 5. edge correlation >= 0.27 6. paint IoU >= 0.96 7. at most 441 strokes 8. both fresh-process replays of the trajectory meet this and end in the same state |
| Family | oil |
| Picture | a square crop of Vincent van Gogh's Cafe Terrace at Night, 1888 |
| Difficulty | extreme |
| Tier | privileged |
| Continuous Score | 1 - mean ΔE / 20 |
| Grader Facts | name — cafeterrace alpha — truth_alpha.npz n_strokes — 294 source — the reference picture: File:Vincent Willem van Gogh - Cafe Terrace at Night (Yorck).jpg (public domain), cropped and downscaled license — Public domain area_mm2 — 51612.8 |
| Scoring | s.score() is served by a separate scoring container: it replays the actions the agent executed since its last reset on the graders' copy and returns the metrics, never images or target data |
| Agent Budget | 7200 s of wall clock per mode |
From our RoboPaint task definitions.
Tags¶
Task DomainManipulation
Why this task is interesting¶
Not yet written.
Capability notes¶
Not yet written.
Oracle demo review¶
Not yet reviewed.