Camera-to-action v0.1: reproducible research candidate

Camera-to-action v0.1 adds a local, reproducible loop from synchronized simulated demonstrations to an installed PiPER motor-policy candidate. It runs through Training Studio or dvidia-train robot, with dvidia-arm for installation and simulation execution.

This candidate did not qualify. The complete registered unseen-start results are:

Controller	Strict placements	Critical-contact episodes
Learned camera/encoder policy	5/64	2/64
Matched blind learner	9/64	34/64
Fixed primitive	31/64	30/64
Authored teacher	64/64	0/64

The learned-minus-fixed difference is −40.625 percentage points (97.5% paired group-bootstrap interval −60.9375 to −20.3125). The learned-minus-blind difference is −6.25 points (−20.3125 to +7.8125). All 16 negative controls behaved as specified. Lower contact counts do not establish a successful skill or physical safety.

The registered study covers 452 scheduled attempts, including failures and refusals, with no replacement starts, teacher fallback in evaluation or confirmation-driven tuning. The completed independent audit reconstructed all 452 saved attempt outcomes (including early refusals) and verified 2,384,532 files with zero mismatches; it reran no physics or model inference. Audit SHA-256: 2774959f8e1a5be7283cea129da0ffea000d8375eb445140c8c7a60ab15efed8. Demonstrations and correction labels come from a privileged simulator teacher. The model uses a historical visual anchor, causal masked camera/encoder history and previous applied commands; native object/contact truth and the current teacher command are excluded from inputs.

The selected corrected model is 543,977 bytes. Its measured fitting time was 10.95 seconds after verified dataset export. Inference p95 was 0.447 ms, while the full offline decision boundary p95 was 36.432 ms on a shared Apple M5 host. These are paused-simulation measurements, not a real-time robot control guarantee.

The main failure is acquiring and carrying the block: 22/64 grasps, 21 lifts, 6 carries and 5 strict placements. Diagnostic tracking returned no object estimate during 749 contact-phase decisions across four teacher recordings. That limitation is preserved, not filled with simulator truth.

Install the wheel below in a fresh Python environment with the documented simulation extras and provision the separately licensed PiPER assets. This prerelease provides the wheel, independent audit, summaries and checksums. The research report contains the protocol, comparisons, limitations and recorded evidence.

Full dataset publication remains incomplete. The separate approximately 36.28 GB frozen dataset/evidence package has not completed publication and anonymous readback on Hugging Face. These release assets do not include that full dataset/model bundle or raw recordings. Upload activity or a dataset placeholder must not be treated as an available, verified archive. The source and five assets below remain independently usable; this prerelease does not claim the Hugging Face archive is available.

Reviewed delivery source: d89b24c82b7534a439bbaa3cac3ef5c9d3b80088. This adds a publishing-only deterministic gate-order fix to ab9082e; all 19 registered runtime files and the installed UI remain unchanged. Registered runtime: c9a7c55531533f60e24651a99d509f721e195051. This is a non-latest prerelease; it does not change stable main or qualify ordinary-video transfer, shoes, other arms or physical hardware.

Source is available at the exact reviewed commit and through GitHub’s generated source downloads. The separately built, SHA-bound source archive is retained for the pending Hugging Face evidence package; it is not attached to this release. Its first redundant GitHub upload was interrupted. The five attached release assets contain the wheel, audit, summaries and checksums.