CLAPBOARDBENCH
Opus 5 and Gemini Robotics ER 2 Preview drive two 6-DoF arms, one motion per LLM call. The task: “Lift the clapper stick, then let go so it snaps shut with a clap.” Each transcript shows every frame the model saw and the note it wrote before each move; watch plays the run back in the Rerun viewer—cameras, notes, and joint traces on one timeline. Outcomes were judged live by a human operator.
| run | model | start utc | steps | length | result | transcript | rerun |
|---|---|---|---|---|---|---|---|
| opus-early | claude-opus-5 | 00:51 | 559 | 17:21 | pass | view | watch |
| opus-1 | claude-opus-5 | 02:18 | 143 | 4:48 | pass | view | watch |
| opus-2 | claude-opus-5 | 02:23 | 188 | 4:39 | fail | view | watch |
| opus-3 | claude-opus-5 | 02:28 | 1235 | 29:04 | fail | view | watch |
| opus-4 | claude-opus-5 | 02:59 | 218 | 8:11 | fail | view | watch |
| opus-5 | claude-opus-5 | 03:09 | 233 | 7:51 | fail | view | watch |
| gemini-1 | gemini-robotics-er-2-preview | 03:20 | 81 | 2:05 | fail | view | watch |
| gemini-2 | gemini-robotics-er-2-preview | 03:23 | 112 | 3:22 | fail | view | watch |
| gemini-3 | gemini-robotics-er-2-preview | 03:27 | 180 | 4:40 | fail | view | watch |
| gemini-4 | gemini-robotics-er-2-preview | 03:31 | 136 | 4:30 | error | view | watch |
| gemini-5 | gemini-robotics-er-2-preview | 03:38 | 55 | 0:49 | fail | view | watch |
Raw EvalLogs, per-trial message transcripts, API wire captures, and Rerun recordings live in the source repo: github.com/robocurve/clapboardbench.