Second-workstation results · 28 September 2026¶
One full RotateValve expert episode and one full DP episode succeeded on a second workstation. This closes the bounded installation/execution check. It does not reproduce training, bitwise trajectories or a 30-state success rate.
| Run | Reset | Success | First crossing | Maximum signed angle |
|---|---|---|---|---|
| Expert | 2500 | Yes | Step 850 | 170.023865° |
| DP | 2500 | Yes | Step 994 | 170.108521° |
| Original DP reference | 2500 | Yes | Step 980 | See original summary |
Both new runs used the original 2240-control-step maximum and exited successfully. The DP threshold crossing is 14 steps, or 0.466667 s at 30 Hz, later than the original record. No training, physics changes, tuning or replacement reset occurred.
Evidence and environment¶
- Returned records, commit
6fcbd80. - Frozen core runtime:
33a35ecb5b35ad6a95df629a395a99cc0fb83d11. - RTX 5090, driver 595.84; Python 3.12.3, Isaac Sim 6.1.0.0, Torch 2.12.0+cu130.
- Separate checkout/environment and simulator/Warp cache; reused download cache and authorized robot meshes. This was not an empty-machine installation.
- The selected DP weights and original evaluator hashes agree with the reference.
- All 29 handoff checksums verified.
paper/build_second_machine.pyindependently recomputes each first signed 170° crossing and checks terminal success/no reset. - The original RGB archive was streamed and verified, including the raw RGB payload; no dataset was retained or used for training by this integration.
Archive SHA-256: 4465c83f6aacaef627e85e699a8e5b0b91c12e0619d28d871d3a8b9efd487060.
Raw RGB SHA-256: aee225e36b83a6669bcabfac332af2aa8bfc3b6813aeb4c7dfb8aefd528f9ad3.
Interpretation¶
The chosen checkpoint can execute successfully in another recorded GPU/software environment on this reset. Timing differs slightly, so no bitwise reproducibility claim follows. Two outcomes do not justify extrapolation to other tasks, model training or aggregate rates. Existing benchmark scores remain unchanged.