PiPER learned placement v0.1 — simulation research

A first learned, simulation-only PiPER placement primitive completes 5/12 frozen unseen-start trials. The authored teacher completes 12/12; open-jaw, hold and shuffled-label learners each complete 0/12. All 60 attempts and the seven failures remain public.

Trained from 32 synthetic state/action demonstrations (32,000 frames), the selected 2,031,879-byte model learns clock-conditioned joint/jaw references with native servos. It has no learned contact feedback, vision-to-action bridge or physical robot qualification.

Downloads: arm-enabled Python wheel, standalone data-only JSON model and complete sealed evidence ZIP. Verify SHA256SUMS.txt before use. See docs/piper-learned-placement.md at this tag for install, scope and reproduction commands. Use --force-reinstall when upgrading an earlier 0.1.0 experimental wheel.

Validation: 36 targeted implementation/driver/scorer tests + 8 publisher tests; 60 independently rescored attempts; 16,016 native replay frames reconstructed with zero object/TCP residual; desktop and 390 px iframe viewport checks; actual fresh-wheel native execution on a disclosed passing development example.

The frozen driver has a documented tuple/list serialization guard issue. The bounded dvidia-piper-confirm adapter resumes persisted, identity-checked tapes without repeating physics; all 60 recovery records are included.

The model and evidence are research artifacts, not a reliable robot skill. Numerical inference p95 is 1.039 ms, while 344/12,000 full control steps exceed 20 ms; sensing, transport and physical actuation are excluded.

Published artifacts: model, demonstrations and complete evidence, recorded 3D replay, and research paper.

Browser hosting fixes are recorded separately from the original sealed study. The model, native traces, scores, installable wheel and release assets remain unchanged. Publication versions and anonymous download verification.