Contract-grounded LLM agents boost robot behavior tree synthesis

A new arXiv paper (arXiv:2607.12220) introduces a contract-grounded architecture for synthesizing robot behavior trees (BTs) from natural language. The approach uses a coding agent that first queries a robot-side Model Context Protocol (MCP) server to retrieve an explicit contract—a skill library, allowed BT operators, and optional composition templates—before generating any BT. This ensures every generated tree only references skills the robot can actually execute, eliminating the brittle reliance on prompt authors knowing robot implementation details.

Evaluated with a closed model (Sonnet 4.6) and a smaller open-source model (Gemma4:31b) across 110 simulated tasks in PyRoboSim and 14 tasks on a physical Husarion Panther robot, the architecture demonstrated near-perfect BT validation and high task success. Notably, BT composition templates significantly improved success on reactive control-flow tasks for the smaller model, and the system transferred seamlessly to physical hardware running a Nav2 stack opaque to both operator and agent. This work provides a practical path for non-experts to deploy robots via natural language while maintaining runtime safety guarantees.