ROS 2 Skill

Agent Skill for ROS 2 robot control via rclpy.

Agent (LLM) -> ros2-skill -> rclpy -> ROS 2 robot

Overview

An AI agent skill that lets agents control ROS 2 robots through natural language. The agent reads SKILL.md, understands available commands, and executes ros2_cli.py to interact with ROS 2 directly via rclpy. No rosbridge required.

The long-term goal is full parity with the ros2 CLI — every command available in a terminal, made accessible to an AI agent. Beyond that baseline, ros2-skill adds capabilities that only make sense in an agent context: goal-conditioned publishing, sensor image capture, system diagnostics, and external integrations like Discord reporting.

Quick Start

ros2-skill is tool-agnostic: any AI agent/runtime that can load and execute Agent Skills can use ros2-skill to control ROS 2 robots.

Common deployment pattern:

• Run the agent session on the robot computer (for example nanobot, Claude, or Claude Code) and load SKILL.md + the rules in references/RULES-*.md
• If needed, connect to that on-robot agent session remotely via your connector/messaging layer; the remote client talks to the agent session, and that session executes ros2-skill locally on the robot

Once configured, you can talk to your robot naturally:

• “What topics are available?”
• “Move the robot forward 1 meter”
• “Trigger the emergency stop”

Agent Workflow: The agent automatically:

1. Understands user intent (subscribe/publish/call/send/run/launch)
2. Uses Path A when a profile is loaded (profile-first static resolution)
3. Uses Path B when profile data is missing or incomplete (live introspection fallback)
4. Uses live introspection for runtime-only state checks in both paths
5. Finds message types and structures
6. Applies safety limits and preflight checks
7. Executes the command

No user clarification needed. The agent uses ros2-skill tools to answer its own discovery and verification questions, including finding and launching the robot stack if it is not already running.

For any deployment, load SKILL.md (command reference) plus the operational rules under references/RULES.md.