DVIDIA Native Action Capture v0.1 preview

A native PiPER capture preview now pairs genuine rendered RGB-D with measured joint/finger state and issued, ramped and applied actuator references. It is an engineering data-collection milestone for one authored simulation trajectory.

• Two 20-second captures preserve 1,000 intervals each; all 1,001 paired RGB/depth frames and recorded motion match exactly.
• A declared interruption retains two complete intervals with an unknown outcome. Missing object-tracker estimates remain null and strict training export refuses.
• 37 focused recorder/capture tests, 19 packaging guards and six viewer contract tests pass. Full and compact viewer browser checks cover playback, scrubbing, separate raw-depth links and interrupted outcomes.
• The 0.1.3.dev1 wheel passes a fresh, hash-locked dependency installation and genuine native capture outside the source checkout. This preview is separate from stable SDK 0.1.2.

Download the wheel for local capture, or extract the evidence ZIP and run python -m http.server 8000 inside it to inspect all original records offline. Install the wheel with its native extra: python -m pip install './dvidia_training-0.1.3.dev1-py3-none-any.whl[arm]'. Fetch pinned assets once using dvidia-robot-sim fetch piper --output robot-assets, then run dvidia-piper-capture run --assets robot-assets --output native-capture --duration 0.2 and dvidia-piper-capture inspect native-capture.

The evidence ZIP is about 106 MB, expands to about 234 MB, and includes the wheel, original PNG/NPY frames, actuator traces, audits, source extracts and license notices. The standalone SDK wheel is about 5.5 MB. Native capture requires Python 3.12+, the optional native dependencies, verified PiPER assets and a working graphics context. macOS arm64/CGL was tested; minimum RAM, other operating systems and headless rendering are not qualified. Once installed and assets are present, capture uses no API key or network.

The CGL renderer explicitly reports limited depth precision. The two full runs repeat one start, and native task-success diagnostics do not qualify a learned skill. The simulator pauses for capture; timing is not a hard real-time guarantee. No object tracker, operator corrective demonstrations, policy training, human-video action recovery or physical-arm transfer is provided.

The first sealed viewer-only revision mistakenly treated raw NPY depth as an image. That local revision remains preserved; the published second revision separates RGB playback from raw-depth downloads. An independent comparison confirms all 10,046 scientific/source/protocol files are unchanged and no physics was rerun for the fix.

Original DVIDIA code/evidence is MIT. PiPER assets retain MIT notices; MuJoCo and other bundled components retain Apache-2.0 and their original notices. Dependencies retain their own licenses. No credentials, private footage or learned weights are included.