AGENT SKILL

Physical AI Video Data Augmentation

暂无简介

License 许可
MIT
所属仓库
rhCat/skillChip
最近更新
2026-07-31
Stars
★ 0
Forks
0

01 前言 Front Matter

name
Physical AI Video Data Augmentation
perks
[ "run", "gen-configs", "prepare-assets", "render-compare" ]
skill
physical-ai-video-data-augmentation

02 资源文件 48

03 SKILL.md 正文

physical-ai-video-data-augmentation — Physical AI Video Data Augmentation

Orchestrate the end-to-end video data augmentation + auto-labeling workflow on OSMO — flow selection, credential preflight, submit-time guard, GPU submit, monitoring, and output retrieval.

Perks

perk tool(s) nature
run preflight_credentials, pre_submit_guard destructive (deploys/repairs in-cluster NIMs, submits GPU workflows to OSMO, writes OSMO credentials)
gen-configs generate_configs read-only inputs (deterministic per-video augmentation configs + manifest under record_store)
prepare-assets prepare_demo_assets pulls the VDA demo dataset from Hugging Face and flattens it into one demo input dir
render-compare render_side_by_side read-only inputs (side-by-side input-vs-augmented MP4 from a staged local run)

The run perk gates a VDA submit: preflight_credentials validates secrets, OSMO credentials, the control plane (ONLINE GPU pool + GPU POD_TEMPLATE), and workflow image access, then pre_submit_guard validates the rendered workflow YAML (setup.files coverage, dataset/cache URLs, VIDEO_NAME basenames) before any osmo workflow submit. Because the workflow deploys NIMs and submits GPU jobs to OSMO, the perk is declared destructive: true; the executor gates it accordingly. Each tool emits one line of structured JSON and writes its report under record_store.

The other three perks are standalone, non-destructive operations that bracket a run: gen-configs runs generate_configs.py to deterministically render one Cosmos augmentation config per input video (plus a manifest.yaml) from a pipeline config dir; prepare-assets runs prepare_demo_assets.sh to pull the nvidia/video-data-augmentation-demo videos from Hugging Face and flatten them into one demo input directory; and render-compare runs render_side_by_side.sh to stitch the original input and the augmented output of a locally-staged run into one side-by-side MP4. Each emits one structured-JSON line and writes under record_store.

How to use it

Copy ledger.json -> task-ledger.json, set the vars (WORKFLOW) + record_store, then validate -> compose -> compile -> oversight -> executor.

Localized from NVIDIA/skills physical-ai-video-data-augmentation (Apache-2.0).

04 同仓库更多 Skills