This is the historical first candidate review. See the current selection and instruction revisions; push_T has been retired.
All three seed-0 attempts finished normally with valid native evidence and complete provider usage. Push T and deposit coin succeeded; nut-on-bolt seating did not, with two of three assemblies complete. Prefer fasten_screws as the hard replacement, keep push_T if contact-mode diversity is the priority, and retain deposit_coin as a secondary candidate pending a small final-state wording clarification. No task is automatically removed by this assessment.
Retain as easier reference
The agent aligned the T block and satisfied the native arm-return tolerance.
Native success after 1,318 action frames, with terminal XY error 5.175 mm and orientation error 0.0586 degrees. The left arm met the native home-pose tolerance during its return; success stopped the episode before exact joint reset. All 71 responses have complete usage and execution finished normally. Image-based planar localization and tabletop motion helpers were saved.
The retained image projection, color/shape localization and push-motion helpers may transfer to new saved layouts. Multi-seed improvement has not been measured.
Instruction alignment: The measured prompt and public describe goal agree. Native conditions check no lift above 1 cm, XY error below 7 mm, orientation error below 7 degrees, and both arms within the native origin-pose tolerance. Return means per-axis end-link error at most 15 cm and orientation error at most 20 degrees, not exact joint reset.
Measured instruction: Push the T-shaped block along the tabletop without lifting it, until its position and orientation align precisely with the gray T-shaped pad. Then return both arms to their starting poses.
Keep as secondary candidate; clarify final-state wording before finalist use
The agent recovered from failed grasps and a failed handoff, then deposited the coin successfully under native criteria.
Native success after 3,246 action frames and 134 model responses, with complete usage and no execution exception. The saved memo records coin slip with a fully closed gripper, a failed dual-arm handoff, recovery from the tabletop, gentler gripping and visual slot alignment. Terminal FK confirms the right end link was within the native home tolerance (largest axis error 42.70 mm; orientation error 6.46 degrees), despite the gripper remaining over the bank. Native success stopped further action.
Saved camera triangulation, TCP-to-flange motion control, gentle pinch settings and recovery notes are plausible reusable learning artifacts. Improvement across seeds has not been tested.
Instruction alignment: The measured instruction is reproduced verbatim in the prompt and describe output. The native predicate checks full coin-bbox containment in the bank between its bottom and center functional heights, plus the broad arm-origin tolerance. It does not require bottom contact, rest, or exact joint reset. The phrase "drop completely to the bottom" is stronger than the measured native condition; use "let it fall fully inside the bank" in a future instruction revision. This published attempt retains its actual measured wording.
Measured instruction: Pick up the coin from the holder and insert it through the slot of the coin bank. Let it drop completely to the bottom inside the bank, then return both arms to their starting poses.
Priority finalist replacement for the removed pour task
The attempt ended normally without native success; two of the three nut-on-bolt assemblies met the native geometry conditions.
The agent stopped after 3,459 of 20,000 available action frames, with native partial reward 0.5, 89 model responses and 88 tool calls. All usage is complete and no execution exception occurred. Independent terminal-pose and asset-metadata recomputation found all three nuts upright and within 1 mm XY of their matching bolt. Red and blue insertion overlaps were 10.747 and 10.332 mm; pink was only 4.497 mm versus a strict greater-than-9-mm requirement. Both arms returned to zero joint positions and both grippers were open. The failure is incomplete pink-nut seating, not budget exhaustion or a provider interruption. A final original head-camera image also contains a blue nut-like residual at an earlier location. This agrees with the agent memo reporting apparent stale images; the sensor cause and its effect on the failure have not been established.
The agent saved localization, approach, transfer and recovery tools, but its final memo explicitly notes that the release-height recipe did not establish full seating. Verifying depth after release is a concrete improvement target for later learning episodes; no such improvement was measured in this seed-0 test.
Instruction alignment: The modified instruction correctly asks to fit upright nuts onto same-color upright bolts, lower them into place, release them and return the arms. Native scoring checks position, tilt, insertion overlap, gripper opening and arm return. It does not check screw rotation, thread engagement or tightening torque. The benchmark measures nut-on-bolt seating.
Measured instruction: Fit each of the three nuts onto the upright bolt of the same color. Keep each nut upright, center its hole on the bolt, and lower it until fully seated. Release all nuts, open both grippers, and return both arms to their starting poses.
Native arm return uses a 15 cm per-axis position tolerance and 20 degree orientation tolerance on the end link. None of these verdicts proves exact joint reset. Videos are native command replays with initial/final state, control, budget and verdict validation; original intermediate states were not recorded.
One seed-0 attempt per task; this is not a cross-seed reliability estimate. Modified instructions and unchanged native predicates are recorded separately. Video receipts state their exact replay validation scope.