---

name: vinn-offline description: "Routes VINN feature caching and non-interactive k-selection workflows for ACT++ BYOL/ResNet episode features." disable-model-invocation: true metadata: disco-role: operating license: MIT

vinn-offline

Use this sub-skill when the task is about caching image features for VINN or choosing a k value from cached episode features.

Typical triggers

• "Cache BYOL features for the simulated cube dataset"
• "Choose k for VINN"
• "How are feature files named?"
• "Why does the raw k-selection script stop in IPython?"

What this sub-skill covers

• Feature caching from per-camera ResNet18 checkpoints.
• Feature-file naming and layout for simulated and cotrain variants.
• Offline nearest-neighbor k selection over cached features.
• CUDA requirements and dataset-index assumptions for VINN preprocessing.

What it excludes

• ACT/CNNMLP/Diffusion training and eval -> policy-training.
• Simulation episode generation / replay / visualization -> simulation-data.
• Real-robot VINN deployment -> root troubleshooting only.

Read these first

• Workflow recipes
• Troubleshooting
• Data formats

Run this helper first

Before a long cache or k-selection job, use check_vinn_stack.py to confirm the repo checkout imports and the CUDA backend is visible.