Skip to content

Commit 8774265

Browse files
author
DVIDIA
committed
Publish reference-model fits and eight failed development placements
1 parent 78ea4fb commit 8774265

17 files changed

Lines changed: 3658 additions & 7 deletions
Lines changed: 21 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,21 @@
1+
MIT License
2+
3+
Copyright (c) 2026 DVIDIA contributors
4+
5+
Permission is hereby granted, free of charge, to any person obtaining a copy
6+
of this software and associated documentation files (the "Software"), to deal
7+
in the Software without restriction, including without limitation the rights
8+
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
9+
copies of the Software, and to permit persons to whom the Software is
10+
furnished to do so, subject to the following conditions:
11+
12+
The above copyright notice and this permission notice shall be included in all
13+
copies or substantial portions of the Software.
14+
15+
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16+
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17+
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
18+
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19+
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
20+
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
21+
SOFTWARE.
Lines changed: 39 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,39 @@
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).
Lines changed: 26 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,26 @@
1+
---
2+
license: mit
3+
language:
4+
- en
5+
tags:
6+
- robotics
7+
- simulation
8+
- demonstrations
9+
- research
10+
---
11+
12+
# DVIDIA PiPER reference training data
13+
14+
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.
15+
16+
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.
17+
18+
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.
19+
20+
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).
21+
22+
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.
23+
24+
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.
25+
26+
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.

0 commit comments

Comments
 (0)