Briefing
Each tree was generated in a separate Claude Code session started with claude -p; Gemma4:31b ran locally through Ollama inside the same Claude Code harness. The agent queries the robot's MCP server for the skill library, tree schema and templates, then sends the tree to the robot's validation gate. On the Panther, the skills wrap a precompiled Nav2 stack (Navigate, Spin, BackUp) plus a YOLO detection node and a heading controller.
In simulation, Sonnet 4.6 reached 90% task success on the core task set against 50% for Gemma, and templates raised Gemma to 83%. On the Panther the two models performed comparably, with success judged by human observation. The University of Texas at Austin authors publish the prompts, contracts and generated trees on a project website.
Written by the Boards & Bots curator