The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill fetches code and model checkpoints from third-party repositories not included in the trusted vendors list.
Clones the primary research repository from https://github.com/moojink/openvla-oft.git (SKILL.md).
Clones the simulation environment from https://github.com/Lifelong-Robot-Learning/LIBERO.git (SKILL.md).
Downloads model checkpoints and adapters from Hugging Face under the moojink user namespace (e.g., moojink/openvla-7b-oft-finetuned-libero-spatial).
[REMOTE_CODE_EXECUTION]: The skill facilitates the execution of scripts and binaries from the downloaded external repositories.
Executes python experiments/robot/libero/run_libero_eval.py and python experiments/robot/aloha/run_aloha_eval.py using content from the cloned repositories.
Runs torchrun with vla-scripts/finetune.py for model training.
Performs editable mode installations (pip install -e .), which triggers the execution of setup.py scripts from the downloaded sources.
[COMMAND_EXECUTION]: The skill uses various system-level commands for environment configuration and task execution.
Utilizes conda create, conda activate, and pip install for environment and dependency management.
Uses git clone to acquire external source code.
Initiates a local API server using uvicorn and fastapi in vla-scripts/deploy.py to serve model actions over the network.
[PROMPT_INJECTION]: The skill processes language instructions to guide robot actions (language conditioning), creating an indirect prompt injection surface.
Ingestion points: Natural language task descriptions (e.g., "pick up the bowl") are processed by the VLA model (SKILL.md).
Capability inventory: The skill has the ability to execute subprocesses (training/eval scripts) and perform file system writes (checkpoints and logs).
Sanitization: No explicit sanitization or boundary markers for language instructions are documented in the skill's workflows.

fine-tuning-openvla-oft