LLM-HBT: Dynamic Behavior Tree Construction for Adaptive Coordination in Heterogeneous Robots

Basic Information

• Paper ID: 2510.09963
• Title: LLM-HBT: Dynamic Behavior Tree Construction for Adaptive Coordination in Heterogeneous Robots
• Authors: Chao-ran Wang, Jingyuan Sun*, Yan-hui Zhang, Mingyu Zhang, Chang-ju Wu*
• Classification: cs.RO (Robotics)
• Publication Date: October 11, 2025 (arXiv preprint)
• Paper Link: https://arxiv.org/abs/2510.09963

Abstract

This paper proposes a novel framework for automatic behavior tree (BT) construction in heterogeneous multi-robot systems, addressing challenges of adaptability and robustness in dynamic environments. Traditional robots are constrained by fixed functional attributes and cannot efficiently reconfigure strategies when tasks fail or environments change. To overcome this limitation, the authors leverage large language models (LLMs) to dynamically generate and extend behavior trees, combining the reasoning and generalization capabilities of LLMs with the modularity and recovery capabilities of BTs. The framework comprises four interconnected modules—task initialization, task assignment, BT update, and failure node detection—forming a closed-loop operation. Robots execute their BTs during operation, and when encountering failure nodes, they can either locally extend the tree or invoke a central virtual coordinator (Alex) to reassign subtasks and synchronize companion BTs.

Research Background and Motivation

Core Problems

1. Insufficient Adaptability: Traditional multi-robot systems struggle to generalize in dynamic and unstructured environments, heavily relying on predefined priors and limited training data
2. Rigid Decision Frameworks: Existing decision frameworks are either too rigid to support online reconfiguration or too fragile to ensure long-term robustness
3. Heterogeneous Coordination Challenges: Heterogeneous robots possess different capabilities; how to collaboratively reconstruct and share behavior trees at runtime remains insufficiently addressed

Research Significance

Multi-robot systems hold tremendous potential for improving operational efficiency, but must adapt to failures, environmental changes, and unexpected situations in dynamic environments. This is critical in practical applications such as search and rescue, warehouse automation, and service robotics.

Limitations of Existing Approaches

1. LLM-based Methods: While demonstrating strong reasoning capabilities, they typically generate task plans in a one-shot manner, lacking online correction mechanisms after execution begins
2. Behavior Tree-based Methods: Provide modularity and recovery mechanisms but heavily depend on manually designed action nodes and predefined task structures
3. Lack of Unified Framework: Existing research fails to adequately integrate LLM's semantic reasoning capabilities with BT's structural robustness

Core Contributions

1. Dynamic Framework: Proposes a dynamic framework integrating large language model reasoning with behavior trees for heterogeneous multi-robot coordination
2. Hybrid Mechanism: Designs a centralized-distributed hybrid mechanism enabling runtime adaptation through local BT extension and centralized task reassignment
3. New Benchmark: Constructs a new benchmark encompassing diverse simulated tasks and real-world environments, validating method robustness and scalability
4. Closed-loop Execution: Implements a closed-loop cycle of failure detection, reasoning, and tree adaptation, enabling heterogeneous robots to continuously optimize execution strategies

Methodology Details

Task Definition

Consider a heterogeneous multi-robot system (HMRS) R = {r₁, ..., rₙ}, where each robot rᵢ possesses an action space:

$A_i = \{a_i^1, ..., a_i^{m_i}\}$

Heterogeneity is manifested in Aᵢ ≠ Aⱼ (i ≠ j), reflecting morphological and capability differences. A task τ is represented by a required action set Aτ ⊆ ⋃ᵢ Aᵢ.

Model Architecture

1. Overall Framework Design

The framework contains four interconnected modules:

• Task Initialization: Converts human instructions into initial BT
• Task Assignment: Triggered by failure nodes to invoke central dispatcher for task reassignment
• BT Update: Inserts new subtrees or synchronizes BTs across robots
• Failure Node Detection: Continuously monitors BT execution and identifies bottlenecks

2. Behavior Tree Formalization

Behavior tree T = (V, E, r), where V is the node set, E defines parent-child edges, and r is the root node. Each node returns Success, Failure, or Running status.

Sequence node execution logic: