Arm interfaces v0.1 — simulation adapters and compatible skill installs

DVIDIA Training 0.1.3.dev3 adds a common local dvidia-arm CLI for PiPER, SO-101 and UR5e/Robotiq native simulation interfaces. Each has explicit units, limits and asset/runtime identity. The existing PiPER policy can be checked, installed and run separately; incompatible arms and hardware backends are rejected.

A data-only task-feasibility gate checks declared reach, grip fit, mass/friction uncertainty, force/load limits, freshness and test-lift status. It does not sense or actuate a robot.

Validation: 54/54 fixed unloaded mechanics attempts numerically valid; 21,654 state rows; 1,260 issued commands; independent audit of 648 RMS/max values with zero mismatches. All 21 offline compute-budget exceedances remain in the evidence. Two deterministic repeats are timing samples, not task generalization.

Fresh installed-wheel smoke passed outside checkout. Exactly one selected PiPER block placement passed; it does not change the original 5/12 held-out result. This run's full control-step p95 was 34.924 ms and 179/1,000 nominal 20 ms compute budgets were exceeded on the shared host. No real-time or physical qualification.

15 adapter tests (including three native builds), 34 installer/feasibility/benchmark/runner tests and 5 guarded-publisher tests passed. Upstream model licenses are preserved.

Install with Python 3.12: create a virtual environment, verify the wheel against SHA256SUMS, then python -m pip install 'dvidia_training-0.1.3.dev3-py3-none-any.whl[arm]'. Run dvidia-arm list and follow the source guide for explicit asset setup.

Guide · Hugging Face evidence · Research

The preview is simulation-only. Cross-arm policy transfer, hardware adapters and live probe/test-lift integration remain future work. The existing stable release is unchanged.