| --- |
| license: mit |
| language: |
| - en |
| tags: |
| - robotics |
| - simulation |
| - mujoco |
| - benchmark |
| - dvidia |
| size_categories: |
| - n<1K |
| --- |
| # DVIDIA arm interfaces v0.1 |
|
|
| Three native CPU simulation interfaces: PiPER, SO-101, and UR5e with Robotiq 2F-85. |
|
|
| 54 fixed unloaded mechanics attempts were numerically valid. This is **not 54 successful manipulation tasks**. All traces are retained: 21,654 state rows, 1,260 commands and **21 offline compute-budget exceedances**. An independent audit reconstructed 648 RMS/max values without a mismatch. Deterministic repeats measure timing variability, not unseen-task generalization. |
|
|
| Arm error is in radians. Each gripper retains its actual command units: PiPER single-finger metres, SO-101 hinge radians, and Robotiq native 0–255. These are not interchangeable jaw gaps or calibrated force readings. Terminal-frame tracking is not calibrated fingertip accuracy. The interactive shared host is not a real-time test rig. |
|
|
| ## Run locally |
|
|
| The 0.1.3.dev3 preview provides `dvidia-arm list`, `describe`, `check`, `install`, `inspect`, `run`, `benchmark`, `task-template` and `assess-task`. |
|
|
| [Install the preview](https://github.com/Dvidia-Inference/dvidia-training/releases/tag/arm-interfaces-v0.1-preview) · [Source and commands](https://github.com/Dvidia-Inference/dvidia-training/blob/feat/arm-interfaces-v01/docs/arm-interfaces-v01.md) · [Research](https://research.dvidia.org/papers/arm-interfaces-v01/) · [Skillspaces](https://dvidia.org/store?view=spaces) |
|
|
| Only the existing PiPER learned block policy has a runner in this installer. Its frozen held-out placement result remains 5/12; this dataset does not train a new model. The policy is rejected on the other arms. Setup needs network access; local simulation uses CPU and verified assets without a CUDA requirement. Hardware minimums and physical transfer remain unqualified. |
|
|
| ## Evidence |
|
|
| - `arm-interfaces-v01-evidence.zip`: every native trace, commands, results, fixed protocol, source snapshots, independent auditor, portable verifier and upstream notices. |
| - `summary.json`: all 27 condition summaries and measurement scope. |
| - `independent-audit.json`: read-only numerical and digest verification. |
| - `installation-smoke.json`: isolated installed-wheel validation; one selected simulation example passed, but is not a new held-out benchmark. Its full control-step compute p95 was 34.924 ms with 179/1,000 nominal 20 ms budgets exceeded on the shared host; no real-time qualification. |
|
|
| Extract the evidence ZIP and run its `verify.py` to check the inventory and numerical audit. `native_step_compute_s` measures the complete observed-step wrapper, including snapshots and reference forward kinematics; it is not isolated physics cost. |
|
|
| The task-feasibility command is a data-only planning assessment. Caller-declared reach, object width, mass/friction uncertainty, force budget, calibration, freshness and test-lift status produce blocked/probe/test-lift/ready-for-simulated-transport decisions. It performs no sensing or motion and is not integrated into the learned controller. |
|
|
| ## Rights and limits |
|
|
| Original DVIDIA tooling and synthetic records are MIT. The archive preserves distinct upstream PiPER MIT, SO-101 Apache-2.0, UR5e BSD-3-Clause and Robotiq BSD-2-Clause notices. No raw human footage, downloaded robot meshes, credentials or new learned weights are included. There is no hardware adapter, physical skill qualification or automatic cross-arm transfer. |
|
|
