Flat spatula · flip over¶
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 DexToolBench 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✗ 1h 00m 0% | 09-29 02:51 | 1h 00m | — | 9.4M / 130k | $0.186 | grader_error 0, missing_trajectory 1 |
| L✗ 53m 0% | 09-29 12:09 | 53m | — | 25.4M / 130k | $0.349 | deterministic 1, grader_error 0, video_rendered 1 |
1 other run(s) of DexToolBench
Codex + GPT-6 Luna, reasoning xhigh (Azure OpenAI API, 2026-09-30 stress test)
| Trial | Started | Agent time | Model requests | Tokens in / out | Est. cost | Grade |
|---|---|---|---|---|---|---|
| U✗ 1h 00m 0% | 09-30 18:55 | 1h 00m | — | 14.8M / 177k | $0.277 | grader_error 0, missing_trajectory 1 |
| L✗ 53m 0% ⓘ | 09-30 14:35 | 53m | — | 19.9M / 124k | $0.280 | deterministic 1, grader_error 0, video_rendered 1 |
Task instruction (upstream)
Pick up the flat spatula and move it through the demonstrated motion (flipping something over): bring it to each of the 45 goal poses in order.
Recorded by us: SimToolReal's pretrained RL policy replayed in our MuJoCo port of the task, four views (front, side / top, oblique); it reaches 7 of 45 goals here. The green ghost tool is the current goal.
What DexToolBench states about this task
| Defined in | dextoolbench/trajectories/spatula/flat_spatula/flip_over.json |
| Category | spatula |
| Object | flat_spatula |
| Task | flip_over |
| N Goals | 45 |
| Tool Model | assets/urdf/dextoolbench/spatula/flat_spatula/flat_spatula.urdf |
| Table | table_narrow_bowl_plate.urdf (a bowl and a plate) |
| Metric | task progress: goals reached / goals (8 grasp-box keypoints within 1.5 cm; 10 s per goal) |
From https://github.com/tylerlum/simtoolreal @ 313d5ae.
Measured on the stack we run — MuJoCo 3.3.7 (CPU, bit-exact replay), KUKA iiwa 14 (MuJoCo Menagerie 1b86ece) + Sharpa HA4, 600 Hz physics, 60 Hz control
| Harness Task | dextoolbench_spatula_flat_spatula_flip_over |
| Oracle | SimToolReal's pretrained policy, recorded in this scene and replayed: 7/45 goals |
Read from run in our MuJoCo port (robot_coding_bench tasks/dextoolbench_spatula_flat_spatula_flip_over, image rcb-mujoco 0.1.1), 2026-09-28.
Tags¶
Why this task is interesting¶
Not yet written.
Capability notes¶
Not yet written.
Oracle demo review¶
Not yet reviewed.