Warehouse Ops Center — Physical AI Safety Reasoning

Cosmos Reason 2 (CR2) is the system: one model, multiple safety reasoning tasks on video, delivered as structured JSON + <think> reasoning traces.

This repo is driven by:

• warehouse-ops-center-spec.md (project spec)
• lessons_learned.md (Nebius/vLLM operational runbook and gotchas)

Requirements

• Python 3.11+
• Nebius-managed vLLM endpoint serving nvidia/Cosmos-Reason2-8B

Setup

python -m venv .venv
.\.venv\Scripts\Activate.ps1
pip install -r requirements.txt
copy .env.example .env

Edit .env with your Nebius endpoint IP and API key.

CLI

Analyze a single video:

python -m src.cli analyze --mode load --video .\data\videos\clip.mp4

Modes: load, safety, security, timeline, full

Run a batch manifest:

python -m src.cli batch --manifest .\batch\batch_manifest_example.yaml

Render the exact request payload (offline; no Nebius calls):

python -m src.cli render --mode load --video .\data\videos\clip.mp4

Parse a saved raw output into JSON + <think> (offline):

python -m src.cli parse --raw .\outputs\clip.mp4__load.raw.txt

Run evaluation (offline, against hand-labeled ground truth):

python -m src.cli eval --results .\outputs --ground-truth .\data\ground_truth --out .\outputs\eval_report.json

Run tests:

python -m unittest discover -s tests

Nebius Notes

• Assume reasoning_content is not available. We parse <think>...</think> from the raw assistant message.
• We do not assume we can change vLLM startup flags on Nebius-managed deployments.

Prompting Convention

We follow the NVIDIA Cosmos reason prompt guide (see warehouse-ops-center-spec.md Section 3B): media-first ordering + the standard reasoning suffix appended to the user prompt.