The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: Fetches model source code, task suites, and datasets from well-known services including GitHub (moojink/openvla-oft, Lifelong-Robot-Learning/LIBERO) and Hugging Face (openvla/modified_libero_rlds). While these are standard for research replication, they involve executing code from individual-maintained repositories.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to process external language instructions and visual data to generate robotic actions. This creates a surface where adversarial instructions embedded in the environment or prompts could potentially influence the agent's behavior.
Ingestion points: Processes language conditioning and multi-camera image streams (front, wrist, etc.) via the VLA backbone.
Boundary markers: No explicit delimiters or instruction-bypass warnings are implemented for the language input.
Capability inventory: The model generates high-dimensional robotic control actions (7-14 DOF) and runs local inference servers.
Sanitization: No input sanitization or validation logic for natural language instructions is provided in the documentation.
[COMMAND_EXECUTION]: The instructions require executing numerous shell commands for environment setup, model training using torchrun, and deploying a local FastAPI server (vla-scripts/deploy.py) to handle action requests.
[METADATA_POISONING]: Automated security scanners have flagged the project's documentation website (openvla-oft.github.io) as blacklisted. Users should exercise caution when visiting the project page or using it as a resource.

fine-tuning-openvla-oft