The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructs users to clone and install third-party research repositories, specifically github.com/moojink/openvla-oft and github.com/Lifelong-Robot-Learning/LIBERO, which are required to access the training and evaluation scripts.
[INDIRECT_PROMPT_INJECTION]: The skill implements Vision-Language-Action (VLA) models that use language conditioning to generate robot actions, creating a surface for indirect prompt injection via untrusted task instructions.
Ingestion points: Language-based task descriptions processed during inference in experiments/robot/libero/run_libero_eval.py and experiments/robot/aloha/run_aloha_eval.py.
Boundary markers: No explicit boundary markers or instruction-guarding delimiters are identified in the workflow documentation.
Capability inventory: The skill includes model training (vla-scripts/finetune.py), deployment of an inference server (vla-scripts/deploy.py), and real-world robot control execution (experiments/robot/aloha/run_aloha_eval.py).
Sanitization: There is no mention of input sanitization or validation for the natural language task instructions.
[COMMAND_EXECUTION]: The workflow requires the execution of shell commands for environment setup and model training, including pip install, conda create, and torchrun for distributed training.

fine-tuning-openvla-oft