DVIDIA Observation Lab v0.1.0 (research alpha)

Our first open-source DVIDIA Observation Lab release turns permitted local videos into sampled image observations and a reviewable timeline. It includes local object/palm inference, editable proposals, append-only review revisions and saved JSON export.

• Research paper
• Installation and input manifest
• Interactive synthetic review demo
• Six-excerpt engineering results and original receipts
• Exact upstream artifacts, licenses and credits

Install Python 3.12+ and FFmpeg/ffprobe, clone this tag, install .[observations], explicitly provision the pinned ~40 MB OpenCV Zoo models, then analyze and review your own permitted videos. Model downloads occur only during the explicit setup step; inference runs locally on a CPU. No GPU, paid inference API or account is required for local execution. Minimum hardware and production throughput remain unmeasured.

The public Space uses two authored geometric clips/annotations to demonstrate the same review interface. It performs no live inference and uploads no footage or review edits; revisions remain in the visitor's browser. It is a workflow demonstration, separate from the real local detector pipeline.

Each recorded pilot run completed six public excerpts (19 seconds, 38 sampled frames). This establishes engineering integration, not annotation accuracy, automatic task segments or a trained robot skill. Outcomes remain unknown unless a reviewer supplies a visual assessment. Automatic action descriptions, validated depth/contact and robot learning remain subsequent experiments.

Publication review corrected the thread description: four OpenCV threads requested, ten reported in frozen metadata; actual concurrency was not measured. Original evidence is preserved. Original DVIDIA code is MIT; the adapted detector code and separately downloaded upstream artifacts retain Apache-2.0 notices. No raw footage, model weights or credentials are included in these release assets.

The source archive includes the model provisioning helper and complete docs; the wheel includes runtime commands, the review HTML and both upstream notices. release-manifest.json binds their bytes to source commit 3cc4a3b03ca6d61479c9f6a86cebf7ccc77fd117. The package version remains 0.1.0; this feature tag identifies the Observation Lab build separately from the older training pilot. This research alpha tag is installable without merging the earlier reaction/motor-memory review branches.

Verification: 311 local repository tests, 22 publisher tests, six synthetic preparation/privacy/media checks, browser-local revision checks and all nine exact-source CI jobs passed, including actual Linux ONNX detector loading/inference. A neutral-image smoke test is not an accuracy benchmark.

v0.1.0 — Local Skillspace training (alpha)

DVIDIA Training v0.1.0 is a standalone, MIT-licensed local CPU pipeline for Skillspace footage. It checks real media, preserves source groups, fits a visual model, and exports inspectable weights and measurements. The optional arm extra fits simulation movement candidates from compatible aligned action labels.

Download the source archive and follow README.md, or install the wheel in a Python 3.12+ virtual environment:

python -m pip install dvidia_training-0.1.0-py3-none-any.whl
dvidia-training-doctor
dvidia-training-studio --directory runs/studio

FFmpeg and ffprobe must be installed separately. The wheel needs NumPy 2.5.3. No GPU, account, API key or hosted inference is required. For disconnected installation, prepare a matching wheelhouse and tools first; see docs/installation.md. This package is not published to PyPI.

Provisional small-example sizing is 1 modern CPU core, 2 GiB RAM and 1 GiB scratch space, plus recordings and installed tools. Recommended headroom is 2–4 cores, 4–8 GiB RAM and at least 2 GiB scratch space scaled to data. These are estimates, not verified minimums.

All eight software checks passed on runtime commit d1f0cbc: visual fitting/studio tests on Linux, macOS and Windows with Python 3.12/3.13, the Linux native arm pilot, and actual Docker training and inspection with networking disabled. Release commit 1718c3b only updates documentation after that tested commit.

The example Skillspaces are explicitly authored synthetic data. The visual result is prediction, not an acquired robot skill. The native example uses simulator action labels and an authored gripper/task supervisor; its candidate passed one exact simulated scene. Neither example learns shoe organization from human video or qualifies physical hardware.

SHA256SUMS covers the attached source, wheel, examples and result archives. Original footage is excluded from result archives; example Skillspace archives supply it separately. See the repository hardware, installation and data-format guides for measurement scope and limits.

Frozen sample-count benchmark · 8 October 2026

A reproducible CPU benchmark for the current DVIDIA Training source checkout: fit nested sets of 2, 4 and 7 connected native recording groups, install their candidate capsules, and evaluate them on a frozen twelve-layout suite with teacher and matched open-jaw controls.

Nominal qualifications: 8/8, 7/8, 8/8. Actuator-stress qualifications: 3/4 for each count. All 36 open-jaw controls fail; no execution errors were discarded. The seven-group candidate records 0.267 mm mean and 0.483 mm maximum nominal target error. The complete 84-episode benchmark takes 102.727 seconds on one Apple M5 CPU host; scopes are specified in the report.

This is one composition order using aligned simulator labels, privileged state and authored phases/gripper logic. It establishes neither a sufficient human-video count nor hardware qualification. Repeating this seed repeats known layouts.

• Results, protocol, candidate payloads and reproduction guide
• Benchmark tool
• Public source dataset

The attached evidence ZIP retains native traces, qualification/control receipts and replay files; duplicated source media and training intermediates are omitted. SHA256: bea158d27b754ea547693330ddbfacb15b905e32ba8f574341ea1ad8eba192ba (44,732,405 bytes).

This prerelease records an experiment and the current source tool. It does not replace the frozen v0.1.0 pilot or publish a new PyPI package. Install the benchmark from this source revision with the optional [arm] dependencies. The runtime and capsule compatibility contract remain v0.1.0.