DVIDIA Training 0.1.2: action demonstration tools

DVIDIA Training 0.1.2 adds dvidia-action-demo to the public default branch: a recorder, validator, learning-target exporter and read-only recording comparison for explicit PiPER motor evidence.

It keeps observations, requested/issued commands, ramped targets, applied controls and measurements distinct. Strict checks cover native channel order/ranges, finger-slide units, calibration bindings, causal frame arrival, stale or missing inputs, interrupted records and artifact hashes. Standalone targets preserve context and failure labels outside the feature allowlist.

203 required tests pass, including 27 new recorder tests. The 18 local HTTP tests required loopback access. The final wheel passes a separate isolated installed-CLI smoke outside the checkout, including refusal to treat the synthetic example as a collected demonstration.

Download the small ZIP for the wheel, guide, example, source snapshots, validation record and checksums. Or install the wheel directly with python -m pip install ./dvidia_training-0.1.2-py3-none-any.whl. The base package needs Python 3.12+ and NumPy; recorder inspection uses only the standard library and no GPU.

This is a data contract and local tool. Live synchronized RGB-D/actuator capture, reviewed corrective demonstrations, policy training and physical qualification remain separate work. The example contains synthetic metadata and no real pixels. Replay-check compares two producer recordings; it does not execute physics or robot commands.

Source: https://github.com/Dvidia-Inference/dvidia-training/tree/training-v0.1.2
Guide: https://github.com/Dvidia-Inference/dvidia-training/blob/training-v0.1.2/docs/action-demonstrations.md
Research: https://research.dvidia.org/

Original DVIDIA code and example: MIT. DVIDIA research · @iammrriver · hello@dvidia.org.