// the find
nasa-jpl/rosa
ROSA 🤖 is an AI Agent designed to interact with ROS1- and ROS2-based robotics systems using natural language queries. ROSA helps robot developers inspect, diagnose, understand, and operate robots.
ROSA is a LangChain-based agent that lets you control and inspect ROS1/ROS2 robots via natural language. It ships as a pip package with built-in tools for topics, nodes, logs, and system state, and is designed to be extended with custom tools for specific robots. Aimed at robotics engineers who spend too much time remembering rostopic/rosnode incantations.
LLM-agnostic design — you bring your own model, so you're not locked into one provider or paying JPL's bill. The tool separation (ros1.py, ros2.py, system.py, log.py) is clean and makes adding robot-specific tools obvious. Backed by a published arXiv paper and real JPL deployments on Spot hardware, so the demo isn't just a toy. Solid repo hygiene: CI, publish pipeline, CHANGELOG, CONTRIBUTING, GOVERNANCE — this is maintained like a real product.
LangChain as a core dependency is a significant adoption risk — the framework has a history of breaking changes between minor versions, and you're inheriting that instability. No mention of safety guardrails: there's nothing stopping the agent from sending a 'go full speed' command to a physical robot if the LLM misunderstands the request. The IsaacSim integration is perpetually 'coming soon' with no timeline. Test coverage appears to mock ROS rather than test against a real ROS environment, so the tests won't catch ROS API drift.