A few weeks ago @jopequ shared mcp-rosbags + mcp-lab ( Another MCP Server to analyze your rosbags with LLMs + a UI to benchmark it against different LLM providers ) — an MCP server with purpose-built analysis tools (trajectories, laser scans, tf, plotting) and a neat UI for benchmarking providers on it. We’ve been working on the same problem at Pixel Robotics (AMRs for pallet transport) and just wrote up our approach, which lands on a different point in the design space, so I thought a comparison might be useful to the community:
- we use mcap-mcp-server, that uses a database under the hood. This allows complex queries / joins to the rosbag for fast analysis. This also means that no specific tools are needed, the LLM already knows how to use SQL
- by extending nav2 BT logging, we can query the behavior tree log efficiently
- we share an agent skill that sets up everything to get started
More details, an example walk through and a video in the blog post! How we analyze robot behavior with LLMs — rosbags as SQL