Physical AI People Attribute Search

Physical AI People Attribute Search Workflow Orchestrator

Default workflow skill for PAS execution on OSMO. It owns flow selection, preflight, submit-time interpolation, monitoring, and output retrieval.

Purpose

Run the PAS image augmentation and auto-labeling pipeline safely and reproducibly from preflight to output download.

The PAS pipeline augments existing person-crop datasets by generating controlled clothing/appearance variations (image-domain) and synonymous attribute captions (text-domain). It uses the paidf-augmentation container for image-edit augmentation with MCQ verification, and the paidf-auto-labeling container for person-attribute captioning.

Do NOT use this skill for container-internal tuning-only questions.

Prerequisites

Confirm these before running preflight or any submit. Missing required secrets surface as USER_INPUT_REQUIRED: from scripts/preflight_credentials.sh.

Instructions

1. Select the workflow (e2e, augmentation, auto_labeling) from user intent.
2. Provide a tentative execution-time overview before starting run actions.
3. Run preflight and readiness checks before submit.
4. Derive submit-time values from the active dataset backend (never guess storage_url).
5. Submit the workflow with explicit interpolation values and monitor to completion.
6. Retrieve outputs and summarize task outcomes.

Use run_script(...) for script execution. Canonical examples:

run_script("bash scripts/preflight_credentials.sh --workflow assets/configs/osmo/e2e.yaml")

Available Scripts

Use script-level --help for exact arguments.

Supported Flows

Pick the right workflow for the user's request

Disambiguation: handle vague requests before committing

Default to autonomy: ask only when missing information blocks execution.

Autonomous defaults (do NOT ask)

• If flow is not explicitly requested, default to e2e.
• If cookbook is not specified, default to default.
• If n_augmentations is not specified, default to 3.
• After any stage completes successfully, continue to the next stage immediately.

Triggers that should pause for disambiguation

Step 0: Select Flow and Gather Inputs

Input data policy

• PAS requires person-crop images organized as <person_id>/<view>.jpg subdirectories.
• Always preserve user-provided dataset inputs as first-class.
• Never replace an explicit user dataset with demo assets.
• If no dataset is provided, ask for one (PAS has no built-in demo dataset).

Collect only missing values:

1. Dataset source (storage_url + dataset name).
2. Flow (e2e, augmentation, auto_labeling); default to e2e.
3. OSMO gpu_platform (auto-select when unambiguous).
4. Endpoint URLs for Image Edit, VLM, and LLM — optional; default to in-cluster NIMs and only set for external endpoints.
5. Number of augmentations per person ID (default: 3).

Generate run stamp before each submit:

STAMP=$(cat /proc/sys/kernel/random/uuid | cut -c1-8)
RUN_ID="run-$STAMP"

Execution Time Overview (required before run)

Before running any mutating command, provide a short ETA overview.

Baseline ranges:

Workflow runtime (depends on dataset size and endpoint latency):

Common Preconditions (all flows)

1. Credential and control-plane preflight

   bash scripts/preflight_credentials.sh --workflow assets/configs/osmo/<flow>.yaml
   

   If output contains USER_INPUT_REQUIRED:, ask one concise unblock question.

2. Storage interpolation policy

   storage_url must be derived from the actual dataset/upload backend. Never silently default to stale values on mismatched backends.

3. Inference policy (non-negotiable)

   • Reuse healthy in-cluster persistent NIM endpoints by default (qwen-image-edit-2511, qwen3-vl, qwen25-14b).
   • If missing/unhealthy, deploy automatically — this is a prerequisite, not a user decision. Do NOT pause to ask. See references/nim/README.md for the image-edit NIMService manifest and the VLM/LLM NIM operator install.
   • PAS does NOT launch inference servers inside the OSMO workflow; workers consume the image_edit_url / vlm_url / llm_url endpoints.
   • External endpoints are opt-in only (explicit request or explicit URLs); only then override the *_url values at submit.
   • Never scale down/delete existing NIMs to free GPUs.

Submit (all flows)

Every flow uses the same submit shape; only the workflow YAML changes.

SKILLS_DIR="$(cd "$(git rev-parse --show-toplevel)/skills/physical-ai-people-attribute-search" && pwd)"
STAMP=$(cat /proc/sys/kernel/random/uuid | cut -c1-8)
osmo workflow submit assets/configs/osmo/<flow>.yaml \
  --pool <pool> \
  --set-string \
    dataset=<dataset> \
    run_id=run-$STAMP \
    storage_url=<backend-prefix> \
    gpu_platform=<gpu-platform> \
    skills_dir="$SKILLS_DIR"

Endpoints default to the in-cluster NIMs (image_edit_url / vlm_url / llm_url); deploy/reuse them per the Inference policy above. Do not pass these unless using external endpoints.

Compatibility note:

• Use exactly one --set-string flag and pass all key/value pairs after it.
• Do not repeat --set/--set-string flags in the same command.

Common optional overrides (append to the same --set-string list):

cookbook=<cookbook_name> \
n_augmentations=<count> \
image_edit_url=<image-edit-endpoint> \
vlm_url=<vlm-endpoint> \
llm_url=<llm-endpoint>

OSMO Monitoring

# Workflow status + task states
osmo workflow query <workflow_id> --format-type json \
  | jq '{status, tasks: [.groups[].tasks[] | {name, status, exit_code}]}'

# Logs for a specific task
osmo workflow logs <workflow_id> --task <task_name> -n 200

# Output retrieval
osmo data list --no-pager <output_url>
osmo data download <output_url> <local_dir>/

For runs expected to exceed two minutes, send heartbeat updates at least every two minutes.

Post-Run Output

After successful completion, the output directory contains:

For augmentation / e2e:

• <person_id>/aug_<n>/output.jpg — augmented multi-pane image
• <person_id>/aug_<n>/output.txt — natural-language caption
• <person_id>/aug_<n>/output_metadata.json — verification results
• dataset/augmented_data.json — structured dataset with attributes and queries
• dataset/augmented_imgs/ — split per-view crops

For auto_labeling:

• caption_<id>/task/open_qa.json — person-attribute captions grouped by question bank

Supporting files

Use these canonical locations:

• Workflows: assets/configs/osmo/*.yaml
• Runtime scripts: scripts/*.sh
• Flow walkthroughs: references/flows/*.md
• Setup and triage: references/setup.md, references/troubleshooting.md
• Images: references/container-images.md
• Cookbook tuning: assets/cookbooks/default/README.md