GR1 · PosttrainPnPNovelFromPlateToBowlSplitA¶
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 RoboCasa-GR1 runs for every task of a run.
Codex + GPT-6 Luna, reasoning xhigh (ChatGPT login)default runclosed
| Trial | Started | Agent time | Model requests | Tokens in / out | Est. cost | Grade |
|---|---|---|---|---|---|---|
| U✗ 1h 00m | 10-01 21:16 | 1h 00m | — | 20.7M / 139k | $0.319 | deterministic 1, grader_error 0, missing_trajectory 0, n_actions 1220, replay_success 0, video_rendered 1 |
| L✗ 58m | 10-01 20:09 | 58m | — | 30.2M / 143k | $0.413 | deterministic 1, grader_error 0, live_success 0, missing_log 0, replay_success 0, video_rendered 1 |
Task instruction (upstream)
Pick the bell pepper from the plate and place it in the bowl.
Playback speed
Recorded by us: our reference solution, replayed by the verifier (the scene camera).
What RoboCasa-GR1 states about this task
| Success criteria | 1. the bell pepper is in the bowl: touching the bowl, with its centre horizontally within 0.7 x the bowl's horizontal radius of the bowl's centre 2. the bell pepper does not touch the counter 3. the bowl is upright (its z axis has a world-z component above 0.8) 4. both hands are clear: each grip site (left and right) is more than 0.25 m from the bell pepper's origin and more than 0.25 m from the bowl's origin (any_gripper_obj_far, th = 0.25) -- release the bell pepper and move both hands away before the end 5. the distractor object that starts on the plate with the bell pepper is not in the bowl. 6. unlimited (privileged): both fresh-process replays of the handed-in trajectory end in the same state, and the check holds on it 7. limited (standard): the one episode (no reset) is recorded by the service and replays to the same state; the check holds on it live and in the replay |
| Family | robocasa-gr1/PosttrainPnPNovelFromPlateToBowlSplitA |
| Robot | Fourier GR1 humanoid: two arms, waist and two six-command Fourier hands; fixed base |
| Category | Pick and place (novel objects) |
| Instance | the initial state of official demonstration episode 0 (demo_1, HDF5/PosttrainPnPNovelFromPlateToBowlSplitA.hdf5); MuJoCo state, model arrays, task state and RNG frozen in instance.npz (SHA-256 aced434c880c…, checked on load) |
| Deliverable | (T, 24) native controller commands, 1 ≤ T ≤ 3000, 20 Hz |
| Reference Solution | The demonstration's own actions (original: 214 actions), executed from the frozen scene by the robot's own controller: 214 actions in solution/oracle.npz. The demonstration is a reference solution: it is not in the image and the agent cannot reach it. |
| Limited Mode | Standard mode (robot as a service, eai-standard/2.1) of robocasa-gr1-plate-to-bowl-i00-privileged: the same frozen instance and success check, served by the sim sidecar (images/robocasa-gr1/standard/server_casa.py, tool module casa_tool), which records the episode, replays it in a fresh simulator and judges it; the verifier grades the sidecar's record in a container of its own. The agent sees the robot's cameras (RGB-D, calibrated), its own joints and end effector, and the instruction. |
| Oracle | full |
| Base Image | ghcr.io/mll-lab-nu/eai-robocasa-gr1:0.1.0 |
| Agent Budget | 3600 s of wall clock per mode |
| Environment Source | https://github.com/robocasa/robocasa-gr1-tabletop-tasks/blob/4840e671596f93ca03651524b9f72ffb1aadfeff/README.md#L114 |
| Task Dirs | robocasa-gr1-plate-to-bowl-i00-privileged, robocasa-gr1-plate-to-bowl-i00-standard |
From https://github.com/robocasa/robocasa-gr1-tabletop-tasks @ 4840e67, as defined in our task definitions @ 8c5594a43.
Tags¶
CapabilityControl & Coordination
Task DomainManipulation
Why this task is interesting¶
Not yet written.
Capability notes¶
Not yet written.
Oracle demo review¶
Not yet reviewed.