Train and evaluate reference-endpoint PiPER models

.gitattributes

@@ -16,6 +16,7 @@ src/dvidia_training/camera_action_capture.py text eol=lf
 src/dvidia_training/robot_profiles.lock.json text eol=lf
 src/dvidia_training/camera_action_cache.py text eol=lf
 src/dvidia_training/camera_action_reference_export.py text eol=lf
+src/dvidia_training/camera_action_reference_model.py text eol=lf
 src/dvidia_training/camera_action_study.py text eol=lf
 src/dvidia_training/piper_rgbd_tracker.py text eol=lf
 src/dvidia_training/action_demonstration.py text eol=lf

.github/workflows/ci.yml

@@ -84,6 +84,8 @@ jobs:
         run: |
           python -B -m unittest tests.test_camera_action_reference_export -v
           python -B -m unittest discover -s benchmarks/camera-action-v02-head-preflight -p test_preflight.py -v
+      - name: Verify reference-model training and inference on synthetic fixtures
+        run: python -B -m unittest tests.test_camera_action_reference_model -v
       - name: Test footage, visual training and loopback studio
         run: python -m unittest tests.test_footage_data tests.test_footage_train tests.test_footage_pipeline tests.test_footage_benchmark tests.test_training_studio tests.test_doctor tests.test_coverage tests.test_episode_tape tests.test_reaction tests.test_reaction_metrics tests.test_reaction_convergence tests.test_reaction_benchmark tests.test_motor_memory tests.test_motor_metrics tests.test_motor_benchmark tests.test_observations tests.test_observation_models tests.test_motion tests.test_motion_benchmark tests.test_transition_model tests.test_transition_pilot tests.test_transition_cli tests.test_embodiment_calibration tests.test_embodiment_cli tests.test_reference_sim -v
 

benchmarks/camera-action-reference-model-integration-v01/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 DVIDIA contributors
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

benchmarks/camera-action-reference-model-integration-v01/README.md

@@ -0,0 +1,39 @@
+# DVIDIA reference-model training and inference integration
+
+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.
+
+- [Absolute candidate card](models/absolute/README.md)
+- [Delta candidate card](models/delta/README.md)
+- [Derived-data card](data/README.md)
+- [Compact integration receipt](evidence/integration-result.json)
+- [Training reference identity](evidence/training-input-identity.json)
+- [Initial native smoke refusal accounting](evidence/initial-smoke-accounting.json)
+- [All three engineering batches](evidence/development-smoke-status.json)
+- [Independent eight-attempt native audit](evidence/native-smoke-v3-audit.json)
+- [Renderer warnings](evidence/renderer-warnings.json)
+- [Strict-state failure diagnosis](evidence/native-smoke-diagnosis.json)
+
+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.
+
+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`.
+
+
+## Inspect a retained candidate
+
+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:
+
+```python
+from dvidia_training import camera_action_reference_model as reference_model
+
+model = reference_model.load_model(
+    "benchmarks/camera-action-reference-model-integration-v01/models/delta/model.json"
+)
+assert not model["task_qualified"]
+assert not model["physical_robot_ready"]
+policy = reference_model.ModelPolicy(model)
+print(model["variant"], model["training"]["samples"])
+```
+
+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.
+
+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).

benchmarks/camera-action-reference-model-integration-v01/data/README.md

@@ -0,0 +1,26 @@
+---
+license: mit
+language:
+- en
+tags:
+- robotics
+- simulation
+- demonstrations
+- research
+---
+
+# DVIDIA PiPER reference training data
+
+This card describes the derived training inputs for both retained reference-model candidates. This data directory contains an identity receipt; the benchmark does not redistribute raw videos, caches or reference arrays. Both candidate model files are included separately.
+
+The data comprises 96,000 control rows from 96 existing simulation episodes in 32 training starting-position groups: 64 authored-teacher demonstrations covering both destinations, followed by 32 previously collected mixed-control corrections. These are synchronized simulated camera/action data, not recovered actions from human video. Development and confirmation episodes are excluded from fitting.
+
+Each parent row contains 269 causal features, the issued-teacher or separately recorded teacher-query target, group ownership and provenance. Each reference export adds two C-order float64 `(N, 7)` arrays: exact previous normal servo references and feasible target endpoints. Six dimensions use radians; the seventh is the single-finger jaw coordinate in metres. Reset references come from the recording reset; later references come from the preceding step's actual normal ramp state. The original 20 updates are applied to teacher labels. For mixed control this is counterfactual normal-reference supervision, not a statement that those labels were executed.
+
+The parent cache was built through the original strict exporter. The reference-export construction verified sealed recording metadata and caller-pinned parent data; cache/reference loads validate derived bytes without re-reading raw RGB-D frames. This integration experiment did not revalidate raw recordings. Array, protocol, stage, source, parent dataset and manifest identities are in [training-input-identity.json](../evidence/training-input-identity.json).
+
+Source: [DVIDIA training at the benchmark commit](https://github.com/Dvidia-Inference/dvidia-training/tree/78ea4fbb4b2733e7738eb425eeecd7cb67a52195). Training uses only training-data normalization. Current teacher actions, future observations and native object/contact truth are excluded from features. Missing visual observations remain masked, and historical estimates retain their age semantics.
+
+These bounded synthetic demonstrations do not establish transfer to shoes, arbitrary videos, other arms or physical hardware. Model training losses and a single training-trajectory replay do not establish generalization. The original dataset ownership and failure accounting are preserved.
+
+DVIDIA-generated derived data and accompanying documentation use [MIT](../LICENSE). Upstream simulator and robot asset licenses remain separate. No raw human footage, credentials or upstream model weights are included in this benchmark.