Lab
Lab
What we corrected, and what we have not answered yet.
Corrections
Closed. Kept on the record.
- ClosedOct 2026
We withdrew a 38% success rate
An earlier slip-induced reorientation study reported 38% held-out success. A per-episode audit of our own code found 15 successes on 40 held-out seeds. Of these, 5 happened at the first step because the target already lay within tolerance. The other 10 happened while the object fell out of the grasp. No episode, under any policy, reoriented a held object. The claim is withdrawn and the testbed was corrected.
Source: Preliminary simulation study, version 2, October 2026 · Report not yet public
Open questions
Not yet answered.
- OpenOct 2026
Can a policy learn on the corrected testbed?
The corrected gravity-pivoting testbed produces grip-force-controlled pivoting. No learned policy has yet been trained on it, so the audit reports the testbed only.
Source: Preliminary simulation study, version 2, October 2026 · Report not yet public
- Open2026
What does capscope do with a live model?
The 77% figure is an oracle check against the benchmark’s ground-truth tool calls. An end-to-end evaluation with a model remains to be done.
Source: capscope
- OpenOct 2026
Does the pivoting guard hold beyond 40 fresh objects?
The guard was designed after we saw test results, so it was confirmed once, on 40 fresh objects. One of three seeds stayed clearly worse with noise-model sensing.
Source: Residual Pivoting
- OpenOct 2026
The trial-count audit is being revised
Version 2 is published. A revision is in progress.
Source: How Many Trials Back a Force-Aware Manipulation Claim?