Towards An LLM-Driven Unified Conversion Framework for BT and FSM in Autonomous Intelligent Systems

Zhang Qi; Yang Shuo; Zhu Zhengqiu; Zhou Peng; Jiao Peng

cs.AISep 24, 2026

Towards An LLM-Driven Unified Conversion Framework for BT and FSM in Autonomous Intelligent Systems

Authors: Zhang Qi, Yang Shuo, Zhu Zhengqiu, Zhou Peng, Jiao Peng

Organizations: College of Systems Engineering, National University of Defense Technology CHANGSHA, HUNAN, China

Abstract

Finite state machine (FSM) and behavior trees (BT) are widely adopted behavioral modeling paradigms for autonomous intelligent systems. While functionally equivalent and inter-convertible in principle, existing transformation methods between FSM and BT face major challenges in preserving behavioral completeness and avoiding model complexity explosion. To overcome these issues, we propose an LLM-driven unified conversion framework that enables automatic, efficient, and semantically consistent transformation between FSM and BT. Specifically, a novel loop execution BT structure is designed for LLM to accurately capture the loop structure in FSM, thereby preserving behavioral completeness. To mitigate the state explosion problem in BT-to-FSM conversion, a depth compression strategy is introduced with LLM prompt to eliminate redundant control nodes, complemented by differentiated hierarchical conversion rules that collectively reduce the number of required sub-FSM. Simulation experiments in multiple autonomous decision-making scenarios demonstrate that the proposed framework enables an accurate and automated bidirectional conversion between FSM and BT. Furthermore, it significantly enhances the scalability and maintainability of generated models compared to traditional approaches, providing a practical solution for behavior model conversion in consumer-grade autonomous intelligent systems such as service robots, game agents, and smart home devices

Figures & tables

Figure 1: LLM-based bidirectional conversion framework for FSM and BT behavioural models.

Figure 2: Basic FSM structure.

Figure 3: LEBT structure.

Figure 4: Prompts for FSM-to-BT Conversion.

Figure 5: Simplified flowchart for sequence BT.

Figure 6: Simplified flowchart for selector BT.

Figure 7: Structure Diagram of Sequence BT Conversion to FSM.

Figure 8: Structure Diagram of Selector BT Conversion to FSM.

Figure 9: Prompts for BT-to-FSM Conversion.

Figure 11: Predator-Prey Game Simulation Platform.

Figure 12: Comparison of Conversion Accuracy Across Scenario.

Figure 13: Comparison of Conversion Accuracy Among Different LLM.

	scenario1	scenario2	scenario3
ChatGPT	$95.70{{}_{\pm 4.21}}$	$96.95{{}_{\pm 3.61}}$	$93.65{{}_{\pm 4.63}}$
DeepSeek	$96.35{{}_{\pm 3.00}}$	$97.80{{}_{\pm 2.36}}$	$90.65{{}_{\pm 5.58}}$
Qwen	$94.10{{}_{\pm 4.26}}$	$94.75{{}_{\pm 4.99}}$	$93.55{{}_{\pm 4.59}}$
ChatGPT (without SCP)	$32.10{{}_{\pm 9.27}}$	$28.70{{}_{\pm 8.16}}$	$31.90{{}_{\pm 9.38}}$
DeepSeek (without SCP)	$32.80{{}_{\pm 9.10}}$	$31.50{{}_{\pm 8.14}}$	$31.00{{}_{\pm 8.00}}$
Qwen (without SCP)	$32.60{{}_{\pm 9.34}}$	$32.45{{}_{\pm 8.14}}$	$31.05{{}_{\pm 8.24}}$

Table 1: Conversion Accuracy with 95

\%

confidence intervals

	scenario1	scenario2	scenario3
Base	$32.10{{}_{\pm 9.27}}$	$28.70{{}_{\pm 8.16}}$	$31.90{{}_{\pm 9.38}}$
Base+M1	$37.05{{}_{\pm 9.02}}$	$32.55{{}_{\pm 8.99}}$	$33.85{{}_{\pm 9.14}}$
Base+M2	$83.40{{}_{\pm 7.08}}$	$86.85{{}_{\pm 6.89}}$	$84.50{{}_{\pm 7.35}}$
Base+M1+M2	$95.70{{}_{\pm 4.21}}$	$96.95{{}_{\pm 3.61}}$	$93.65{{}_{\pm 4.63}}$

Table 2: Ablation Experiments

Figure 14: BT behavioral model of robot.

	LLMDUCF	NLLM	BSD	CHDS-BT
Reachability	✓	✓	✓	✓
Completeness	✓	x	✓	✓
No Additional Behavior	✓	x	✓	✓
Deadlock Absence	✓	✓	✓	✓
Termination Consistency	✓	x	✓	✓

Table 3: BT to FSM Behavioral Consistency

Figure 15: FSM generated by proposed method.

Figure 16: FSM generated by BSD/CHDS-BT.

Figure 17: FSM generated by NLLM.

	LLMDUCF	NLLM	BSD/CHDS-BT
Computational Complexity	$\text{O}(1)$	$\text{O}(\textit{n})$	$\text{O}(1)$
Graphical Elements	$4\text{{M}+2{k}+1}$	/	$\approx 36\textit{M}$
Active Elements	$4\text{{M}+2{k}+1}$	/	$\approx 29\textit{M}$

Table 4: Comparison of Structural Complexity

	LLMDUCF	NLLM	BSD/CHDS-BT
Tuck Arm subtree	4	4	12
Safe-Move- To behavior	4	4	4
Dock subtree	4	4	17

Table 5: Edit Distance of FSM

Figure 18: Comparison of Conversion Accuracy Across Scenario.

	scenario1	scenario2	scenario3
ChatGPT	$96.95{{}_{\pm 3.22}}$	$98.35{{}_{\pm 1.77}}$	$95.50{{}_{\pm 4.37}}$
DeepSeek	$96.10{{}_{\pm 3.10}}$	$97.95{{}_{\pm 2.82}}$	$94.05{{}_{\pm 4.61}}$
Qwen	$96.30{{}_{\pm 3.93}}$	$96.50{{}_{\pm 3.09}}$	$95.70{{}_{\pm 3.26}}$
ChatGPT (without SCP)	$31.20{{}_{\pm 8.94}}$	$38.70{{}_{\pm 9.15}}$	$33.20{{}_{\pm 9.35}}$
DeepSeek (without SCP)	$31.35{{}_{\pm 9.33}}$	$32.20{{}_{\pm 9.55}}$	$33.90{{}_{\pm 10.65}}$
Qwen (without SCP)	$32.70{{}_{\pm 9.10}}$	$35.60{{}_{\pm 9.24}}$	$32.10{{}_{\pm 9.30}}$

Table 6: Conversion Accuracy with 95

\%

confidence intervals

Figure 19: Comparison of Conversion Accuracy Among Different LLM.

	scenario1	scenario2	scenario3
Base	$31.20{{}_{\pm 8.94}}$	$38.70{{}_{\pm 9.15}}$	$33.20{{}_{\pm 9.35}}$
Base+M1	$33.80{{}_{\pm 9.74}}$	$38.65{{}_{\pm 10.03}}$	$36.05{{}_{\pm 10.96}}$
Base+M2	$34.30{{}_{\pm 9.37}}$	$46.20{{}_{\pm 10.86}}$	$41.55{{}_{\pm 9.69}}$
Base+M1+M2	$96.95{{}_{\pm 3.22}}$	$98.35{{}_{\pm 1.77}}$	$95.50{{}_{\pm 4.37}}$

Table 7: Ablation Experiments

Figure 20: FSM behavioral model of predator.

Figure 21: BT model generated using proposed method.

	LLMDUCF	NLLM	BSD	CHDSs-BT
Reachability	✓	x	✓	✓
Completeness	✓	x	✓	✓
No Additional Behavior	✓	x	✓	✓
Deadlock Absence	✓	✓	✓	✓
Termination Consistency	✓	x	✓	✓

Table 8: FSM to BT Behavioral Consistency

Figure 22: BT generated by BSD.

Figure 23: BT generated by CHDS–BT.

Figure 24: BT generated by NLLM.

	LLMDUCF	NLLM	BSD	CHDS-BT
Computational Complexity	$\text{O}(1)$	$\text{O}(1)$	$\text{O}(n)$	$\text{O}(n^{2})$
Graphical Elements	$\approx 7\textit{M-1}$	/	$8\textit{M+1}$	$8\textit{M+1+6k}$
Active Elements	$\approx 3.5\textit{M}$	/	$4\textit{M+1}$	$4\textit{M+1+3k}$

Table 9: Comparison of Structural Complexity

	LLMDUCF	NLLM	BSD	CHDS-BT
Fire subtree	6	6	$8+\textit{k}$	$8+6\textit{k}$
Gliding behavior	2	2	2	2
Loading subtree	8	8	$10+\textit{k}$	$10+6\textit{k}$

Table 10: Edit Distance of BT

Figure 25: Different cases of node addition in a BT. The added nodes are highlighted in red.

	BT to FSM	FSM to BT
GPT(gpt-4-turbo)	93.85 $\%$	96.20 $\%$
GPT(gpt-4o)	94.55 $\%$	97.05 $\%$
DeepSeek(v3.1)	94.25 $\%$	95.30 $\%$
DeepSeek(v3.2)	95.10 $\%$	96.15 $\%$
Qwen(v2)	92.90 $\%$	95.45 $\%$
Qwen(v3)	93.80 $\%$	95.75 $\%$

Table 11: Comparison of Version Sensitivity

	Failure Cases	Total Retries	Average Retries
BT-FSM(ChatGPT)	27	42	1.56
BT-FSM(DeepSeek)	31	69	2.23
BT-FSM(Qwen)	35	76	2.17
FSM-BT(ChatGPT)	18	41	2.28
FSM-BT(DeepSeek)	24	53	2.21
FSM-BT(Qwen)	23	51	2.22

Table 12: Retry Statistics of Failed Conversion Cases

	LLMDUCF	NLLM	BSD	CHDS-BT
Text File	✓	✓	✓	✓
Image	✓	✓	x	x
Table	✓	✓	x	x
Unseen Encoding	✓	✓	x	x
New Conventions	✓	✓	x	x

Table 13: Input Representation Generalizability

Explore similar work

A Generative Partially Specified Finite State Machine Approach to Complex Behaviour Planning
Jul 17, 2026Kalana Ratnayake, Michael Pritchard, David Hinwood +2Classical Planning Behavior Tree
University of Canberra · Collaborative Robotics Laboratory, Faculty of Science and Technology, University of Canberra, ACT, Australia · RMIT University +1
CABTO: Context-Aware Behavior Tree Grounding for Robot Manipulation
Mar 17, 2026Yishuai Cai, Xinglin Chen, Yunxin Mao +6Behavior Tree Robot Planning
National University of Defense Technology · PsiBot · PKU-Psibot Lab +1
Multimodal Behavior Tree Generation: A Small Vision-Language Model for Robot Task Planning
Mar 6, 2026Riccardo Andrea Izzo, Cristiano Battistini, Gianluca Bardaro +1Large-Scale Robot Demonstration Datasets Behavior Tree
Department of Electronics, Information, and Bioengineering, Politecnico di Milano, Milan, Italy

Jul 17, 2026, cs.ROJ/K move · Enter open · S save

A Generative Partially Specified Finite State Machine Approach to Complex Behaviour Planning

Kalana Ratnayake, Michael Pritchard, David Hinwood, Maleen Jayasuriya+1

University of Canberra · Collaborative Robotics Laboratory, Faculty of Science and Technology, University of Canberra, ACT, Australia · RMIT University+1

Classical Planning Behavior Tree Automata Agent Behavior Modeling

Open paper AskPDF arXiv

Autonomous robots operating in dynamic environments require behaviour planning systems that combine reactivity, interpretability, and adaptability. While Large Language Models have been successfully integrated with Behaviour Trees for dynamic replanning, Finite State Machines, despite their widespread adoption and computational efficiency, remain unexplored for generative approaches. We propose a Generative Partially Specified Finite State Machine (GPSFSM) neurosymbolic architecture that utilises the symbolic and semantic structure of FSMs, including states and event-triggered transitions, to implement Behaviour Planning. This paper introduces the first GPSFSM framework for robotics, featuring Fabric, an FSM engine that parses, validates, and executes behaviour plans that contain Sequential, Recovery, Parallel-Any, and Parallel-All control structures. We extend the Capabilities2 package in ROS2 with an asynchronous event system for behaviour chaining and runtime parameter injection for configurable execution, addressing the ad-hoc function representations that limit current generative systems. PromptTools provides a unified ROS 2 interface to local and cloud LLMs, with prompt buffering, enabling dynamic asynchronous composition of task and context information. Together, these components enable standardised semantic capability descriptions for robot-agnostic development. Experimental evaluation on navigation tasks demonstrates that our GPSFSM approach achieves consistently higher plan-generation success rates than the state-of-the-art BTGenBot system, particularly excelling in zero-shot scenarios where BTs typically struggle, while maintaining comparable or lower planning latency to frontier LLMs. We also demonstrate that our system can generate complex behaviours. We release an open-source ROS2 stack that makes generative FSM planning practical and reproducible for robotic systems.

Abstract

Explore similar work

A Generative Partially Specified Finite State Machine Approach to Complex Behaviour Planning

CABTO: Context-Aware Behavior Tree Grounding for Robot Manipulation

Multimodal Behavior Tree Generation: A Small Vision-Language Model for Robot Task Planning