Contract-Grounded Behavior Tree Synthesis via Coding Agents

2026-07-13

AI Summary

This work addresses the fragility of deploying behavior trees synthesized by current natural language generation methods, which often lack effective constraints grounded in the robot’s actual executable skills. To resolve this, the authors propose a contract-driven behavior tree synthesis framework that retrieves explicit contracts—including skill libraries, valid operators, and compositional templates—from an on-robot Model Context Protocol (MCP) server during agent querying. These contracts impose hard constraints prior to generation, while runtime verification gates ensure execution correctness. By shifting grounding responsibility from user prompts to the system itself, the approach enables non-expert users to issue high-level commands without knowledge of low-level implementation details. Experiments demonstrate near-perfect behavior tree validation and high task success rates across 110 simulated and 14 physical tasks. Moreover, combining small language models with compositional templates significantly enhances reactive control flow performance and enables successful transfer to real-world robotic platforms.

Abstract

Synthesizing deployable robot behavior trees (BTs) from natural language (NL) requires grounding to ensure every generated BT references only skills a robot can actually execute. Existing LLM-based BT synthesis approaches often place this grounding responsibility on the prompt author. This makes deployment brittle when the author does not know which skills the robot can execute, how those skills are parameterized, or how the robot runtime software constrains valid BT structure. This paper proposes a contract-grounded BT synthesis architecture in which a coding agent queries a robot-side Model Context Protocol (MCP) server to retrieve an explicit contract consisting of a skill library, permitted BT operators, and optional BT composition templates, before synthesizing a BT for validation and execution. In our framework, non-expert operators issue NL commands without knowledge of robot implementation details, while a robot runtime validation gate enforces correctness before execution. We evaluate two LLMs, 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. Results show that contract grounding enables near-perfect BT validation and high task success, that BT composition templates substantially recover success on reactive control-flow tasks for the smaller model, and that the architecture transfers to physical hardware running a Nav2 stack opaque to both operator and agent.

Problem

• behavior tree synthesis
• natural language grounding
• robot skill execution
• contract-based validation
• deployable robot behavior

Innovation

• contract-grounded synthesis
• behavior tree
• coding agent
• Model Context Protocol
• robot grounding