PiPER learned grip feedback v0.2 — measured simulation research

A local, CPU-only learned jaw adapter around the frozen PiPER movement primitive, with an installable CLI, recorded 3D replay and complete synthetic evidence.

On 120 frozen simulation attempts, the jaw learner completed 6/24, its matched blind learner 8/24, the original movement model 5/24, the authored teacher 24/24, and the open-jaw control 0/24. All attempts were valid. Live feedback did not demonstrate improved reliability; interrupted feedback completed 0/12.

An independent numeric audit rescored all 360 collected attempts with no native-metric mismatches. The release retains all five failed residual candidates, 32 mixed-policy attempts, all development cases and the complete confirmation matrix. Named-contact diagnostics identify finger-table contact in 12 feedback trials; measured jaw pulses primarily affect acquisition before established grip.

Inference p95 was 0.501 ms and full-step p95 5.648 ms, with 34/24,000 steps exceeding 20 ms. This is synthetic privileged-state simulation research, with no hardware or arbitrary human-video learning qualification.

• Research, limits and replay
• Model and contracts
• Complete 1.8 GB evidence dataset
• Local installation and CLI

The GitHub kit contains the models, installer, compact results and checksums; the complete native/training/development evidence lives in the dataset. Original DVIDIA code/artifacts are MIT; upstream robot/rendering components retain their included notices.