The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides instructions for running a command-line tool (physicalai run) to execute AI policies on a robot.
[DYNAMIC_EXECUTION]: The workflow involves dynamically instantiating objects and loading plugins (e.g., SO101, UVCCamera) via configuration files using the class_path and init_args patterns.
[INDIRECT_PROMPT_INJECTION]: The skill processes external data in the form of YAML configuration files and exported model policies, which serves as a potential attack surface for indirect instructions.
Ingestion points: runtime.yaml configuration file, calibration.json file, and policy model exports located in ./exports/act_policy.
Boundary markers: The instructions rely on the user manually validating configuration fields against a provided schema but do not specify automated sanitization for the inputs.
Capability inventory: The skill facilitates file system reads, hardware I/O via serial ports (/dev/ttyACM0) and video devices (/dev/video0), and command execution through the physicalai CLI.
Sanitization: The instructions recommend a validation loop using fake robots and dry runs to ensure safety before interacting with physical hardware.

pai-run-policy