PiPER grip clock v0.3
A compact learned progress clock chooses when to hold or advance a frozen PiPER movement reference while acquiring a grasp on one 40 g rigid block. The clock uses privileged simulator state and the original arm and jaw reference. Waiting delays both arm progress and jaw release. It does not learn grasp geometry.
The matched blind clock uses the same teacher rows, fit recipe, and tree depth as the selected full clock, with live grip and history inputs masked. The selected pair is full4 and blind4 at matched depth. All four candidate fits, receipts, protocol, freeze, report, independent audit, runtime source snapshots, and integrity hashes are included.
Frozen confirmation results
Strict successes and valid attempts are reported for every scheduled controller. The denominator includes all 128 cases per controller.
| Controller | Strict successes | Valid attempts |
|---|---|---|
| clock | 44/128 | 128/128 |
| blind_clock | 26/128 | 128/128 |
| fixed | 72/128 | 128/128 |
| teacher | 67/128 | 128/128 |
Full minus blind_clock: +14.1 percentage points (approximate 97.5% interval -8.5 to +36.6). Full minus fixed: -21.9 percentage points (approximate 97.5% interval -36.8 to -7.0). These are paired XY-group t intervals with Bonferroni simultaneous 95% coverage under their assumptions. They are exploratory, not a powered qualification. All predeclared follow-up gates passed: false. Hardware qualification remains false.
The independent numeric audit covered 768 attempts (64 training, 192 development, 512 confirmation) with zero mismatches. The confirmation roster contains 32 fresh XY groups, each paired across two known destinations and nominal/post-grip conditions.
Scope
The acquisition teacher checks bilateral pad contact and force plus tool-relative drift over six sampled observations spanning 100 ms. Its latch releases after acquisition. This is acquisition-only supervision, not continuous learned slip recovery or physical grasp certification. The experiment uses one synthetic rigid 40 g block, two known destinations, privileged simulator sensing, and a narrow PiPER starting envelope. It does not establish arbitrary-object, camera-based, other-arm, or hardware transfer.
Shared-host timing describes four spawn-isolated simulator workers on an Apple M5 host. It excludes cameras, transport, rendering, and physical actuation; it is not an isolated throughput result or a hard-real-time guarantee. physical_robot_ready and qualified_robot_skill remain false.
Files and reproduction
This model artifact is published as Dvidia/piper-grip-clock-v03. The complete synthetic demonstration and attempt lineage is in the DVIDIA dataset repository. The reproduction guide and host record are in docs/. Open the recorded 3D comparison, or read the research paper.
Code: DVIDIA-Inference/dvidia-training, release tag piper-grip-clock-v0.3. Contact: hello@dvidia.org · X: @iammrriver.
DVIDIA-authored code, models, and synthetic numeric evidence are MIT licensed. Upstream PiPER geometry and renderer notices remain with their assets and replay; no robot assets or wheel are bundled here.