|
| 1 | +# DVIDIA reference-model training and inference integration |
| 2 | + |
| 3 | +Two small PiPER research candidates now train and serialize successfully from the same 96,000 training rows. The absolute-reference fit took 5.230713 s and the prior-reference delta fit took 5.097628 s on the measured shared CPU host. Both saved models reproduced all 1,000 outputs in the recorded-trajectory replay. This validates the learning/inference bridge, not task completion. |
| 4 | + |
| 5 | +- [Absolute candidate card](models/absolute/README.md) |
| 6 | +- [Delta candidate card](models/delta/README.md) |
| 7 | +- [Derived-data card](data/README.md) |
| 8 | +- [Compact integration receipt](evidence/integration-result.json) |
| 9 | +- [Training reference identity](evidence/training-input-identity.json) |
| 10 | +- [Initial native smoke refusal accounting](evidence/initial-smoke-accounting.json) |
| 11 | +- [All three engineering batches](evidence/development-smoke-status.json) |
| 12 | +- [Independent eight-attempt native audit](evidence/native-smoke-v3-audit.json) |
| 13 | +- [Renderer warnings](evidence/renderer-warnings.json) |
| 14 | +- [Strict-state failure diagnosis](evidence/native-smoke-diagnosis.json) |
| 15 | + |
| 16 | +The third separately registered native batch completed eight attempts: **0/4 strict placements for each candidate**. Cabs made bilateral contact in 3/4 and a qualified lift in 1/4; Cdelta contacted in 4/4 and lifted in 0/4. Neither completed carry. Unexpected contacts occurred in 2/4 Cabs episodes and 4/4 Cdelta episodes. All four delta traces crossed the strict scorer’s finite TCP coordinate envelope of ±0.5 m; this was not a NaN or physics-warning failure. Full per-case evidence is retained, together with the sixteen earlier infrastructure refusals. No candidate was selected or qualified; the proposed 164-attempt screen and frozen v0.1 result are unchanged. |
| 17 | + |
| 18 | +The two candidate model files are included under `models/absolute/model.json` and `models/delta/model.json`. They are usable through the pinned Python API; this new model format is not supported by the qualified-skill CLI installer. Raw training recordings and derived reference arrays are not redistributed here. Original artifacts and source commit `78ea4fbb4b2733e7738eb425eeecd7cb67a52195` are unchanged. Exact original result SHA-256: `6d44d98488ef26f949412aadc94c32174bd5c4ce879442191b9784deff5b325f`. |
| 19 | + |
| 20 | + |
| 21 | +## Inspect a retained candidate |
| 22 | + |
| 23 | +Use a source checkout containing this benchmark and the unchanged runtime bytes identified below. Install the package and its existing NumPy dependency with `python -m pip install -e .`, then run Python from the repository root: |
| 24 | + |
| 25 | +```python |
| 26 | +from dvidia_training import camera_action_reference_model as reference_model |
| 27 | + |
| 28 | +model = reference_model.load_model( |
| 29 | + "benchmarks/camera-action-reference-model-integration-v01/models/delta/model.json" |
| 30 | +) |
| 31 | +assert not model["task_qualified"] |
| 32 | +assert not model["physical_robot_ready"] |
| 33 | +policy = reference_model.ModelPolicy(model) |
| 34 | +print(model["variant"], model["training"]["samples"]) |
| 35 | +``` |
| 36 | + |
| 37 | +Loading does not start a simulator, read a camera or command hardware. Calling `policy(context)` requires the recorder's exact causal context and retained float64 previous reference, beginning at the registered reset; an arbitrary image or joint vector does not satisfy that contract. Any next native experiment requires its own registration; these two candidates remain unchanged. Do not infer operational task success from successful loading or a training-trajectory replay. |
| 38 | + |
| 39 | +Source at measurement: [`78ea4fbb4b2733e7738eb425eeecd7cb67a52195`](https://github.com/Dvidia-Inference/dvidia-training/tree/78ea4fbb4b2733e7738eb425eeecd7cb67a52195). See [the model bridge contract](../../docs/camera-action-reference-model-v01.md) and [integration notes](../../docs/camera-action-reference-model-integration-v01.md). |
0 commit comments