Contract-Grounded Behavior Tree Synthesis
via Coding Agents

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. Generated BTs, contract snapshots, and per-task execution reports are available below.

Method

A coding agent queries a robot-side MCP server for a machine-readable contract, synthesizes a BT, and submits it for validation before execution. The robot runtime retains authority over skill implementation, BT instantiation, and execution.

Configurations

We compare two contract-grounded configurations to isolate the contribution of reusable BT composition priors (rootstocks). Both use the same MCP-mediated synthesis pipeline and the same validation gate.

MCP Interface

The coding agent interacts with the robot exclusively through five MCP tools. It has no access to backend source code, low-level robot APIs, ROS nodes, or execution internals — only the typed interface below.

Results

All numbers are from the paper. Valid@1 measures whether the first-attempt BT passes schema and contract checks. Success measures task completion after execution.

PyRoboSim — Primary Comparison

Valid@1 and Success counts are over all N tasks in the suite (missing BTs count as failures).

Physical Robot (Panther 14)

✓ denotes a perfect score; fractions indicate partial results. V = Valid@1, S = Success.

BibTeX

@misc{salfity2026contractgroundedbehaviortreesynthesis,
      title={Contract-Grounded Behavior Tree Synthesis via Coding Agents},
      author={Jonathan Salfity and Robert Blake Anderson and Mitch Pryor},
      year={2026},
      eprint={2607.12220},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2607.12220},
}