Meta-World · Handle press side¶
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 MetaWorld+ runs for every task of a run.
Codex + GPT-6 Luna, reasoning xhigh (ChatGPT login)default runclosed
Not in this run: not in the sample (robot_coding_bench PR
Task instruction (upstream)
Approach the handle from the side and press it downward.
Playback speed
Recorded by us: our reference solution, replayed by the verifier (the front camera).
What MetaWorld+ states about this task
| Success criteria | 1. The verifier replays your actions from the frozen start in a separate sandbox (twice, in fresh processes; MuJoCo on the CPU is bit-exact) and applies Meta-World's own success check for handle-press-side-v3 (evaluate_state) to the state after the last action. A success reached earlier and then lost does not count: end the trajectory while the task is achieved. 2. unlimited (privileged): both fresh-process replays of the handed-in trajectory end in the same state, and the check holds on it 3. 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 | metaworld/handle-press-side |
| Robot | Rethink Sawyer: a 7-DoF arm whose hand is a mocap body (the wrist does not rotate), two-finger gripper |
| Category | Meta-World v3 |
| Instance | Meta-World's reset randomization with seed 0; MuJoCo state, model arrays, env caches and RNG frozen in instance.npz (SHA-256 1f01512bb2d6…, checked on load) |
| Deliverable | (T, 4) native actions [dx, dy, dz, grip], 1 ≤ T ≤ 500 (Meta-World's max_path_length) |
| Reference Solution | Meta-World's scripted policy for handle-press-side-v3 (metaworld/policies/, removed from the image), run on this instance up to its first success: 48 actions in solution/oracle.npz (recorded by scripts/metaworldplus/freeze.sh). |
| Limited Mode | Standard-mode twin of metaworldplus-handle-press-side-i00-privileged: the same instance and success check (Approach the handle from the side and press it downward.), served as a robot service (images/metaworld/standard/server_mw.py, protocol eai-standard/2.1, tool module mw_tool). The agent sees camera RGB-D, the robot's own state and the target point; no object poses, no reset; one episode. |
| Oracle | full |
| Base Image | ghcr.io/mll-lab-nu/eai-metaworld:0.1.0 |
| Agent Budget | 3600 s of wall clock per mode |
| Task Dirs | metaworldplus-handle-press-side-i00-privileged, metaworldplus-handle-press-side-i00-standard |
From https://github.com/Farama-Foundation/Metaworld @ 59fc34d (Meta-World v3), 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.