DVIDIA task environments v0.2 · local native preview

Build a reusable PiPER practice environment locally: metric object drag controls, an open bin, one to four rigid objects, explicit OpenUSD and CPU MuJoCo physics. The editor and dvidia-environment Terminal commands share the same recipe.

The frozen environment/verifier benchmark accepts 6/6 passive completion fixtures, rejects 6/6 incomplete controls, and rejects 4/4 invalid layouts before simulation. An independent read-only auditor reconstructs 13,262 observations and 1,337 replay frames with zero mismatches. All inputs, attempts, source snapshots and licenses are retained in the 7.1 MB kit. Identical robot geometry is stored once by content hash.

This is native preview 0.1.3.dev2. A fresh Python 3.12 dependency installation passed init, validate, author, run and export outside the checkout, with both objects completing a separate passive settling fixture. Previously verified PiPER assets were reused; fresh asset downloading and other hosts were not qualified.

Download the wheel, then:

python3.12 -m venv .venv
source .venv/bin/activate
python -m pip install './dvidia_training-0.1.3.dev2-py3-none-any.whl[usd]'
dvidia-robot-sim fetch piper --output ./robot-assets
dvidia-environment studio --assets ./robot-assets --output ./local-scenes

Open the displayed loopback address. For the recorded kit, extract it and run python -m http.server 8000 --bind 127.0.0.1 in the extracted folder. The browser edits drafts and replays saved native poses; new physics runs in the local Python process.

Recorded checks · Complete Hugging Face evidence · Research · Terminal guide

The arm holds its home controls. Shoes are original rigid three-box surrogates with assumed mechanics. The benchmark measures environment construction and false-completion rejection; it does not demonstrate learned placements, arbitrary USD import, camera-based control or physical robot transfer. The full recording run takes 16.55 wall seconds for 13.25 simulated seconds on a shared Apple arm64 host, including compilation and evidence export. It is not an isolated physics throughput or real-time guarantee.

Original DVIDIA code/scenes: MIT. PiPER assets retain MIT notices; dependencies retain their own licenses. No new model was trained. DVIDIA research · @iammrriver · hello@dvidia.org.

A separate reader recovery update clears stale results after download failures, offers explicit retry and caches SHA-256-verified shared geometry (64 MiB cap). It is published in a new Space namespace; the original 175-file benchmark kit, wheel and frozen source tag above remain unchanged. Four URL tests and six recovery/cache tests pass, and a browser fixture confirms three failed reads clear prior results before explicit retry recovers the selected case.