Contract-Grounded Behavior Tree Synthesis via Coding Agents

Jul 15, 2026

Researchers introduce a contract-grounded architecture for synthesizing robot behavior trees from natural language commands. Their system uses a coding agent that queries a robot-side server for an explicit contract detailing available skills and constraints, ensuring generated behavior trees are executable and valid. Evaluated on 110 simulated and 14 physical robot tasks, the approach achieves near-perfect validation and high task success rates with both closed and open-source large language models. The architecture also demonstrates transferability to physical robots with opaque runtime stacks.

Why it matters: This work advances the deployment of robot behaviors from natural language by non-experts, improving reliability and safety through explicit contract-based grounding.

Full story at: arXiv Robotics ↗