.png){kind=icon}
.png){kind=icon}
Return
Environment-Driven and LLM-Guided Multi-Robot Task Inference and Allocation Under Temporal Logic Specifications
.png){kind=icon}
PRE
AI
DOI：10.1109/TASE.2026.3659055.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
In multi-robot systems, the successful execution of tasks typically depends on predefined instructions. However, existing approaches encounter substantial challenges in dynamic environments, particularly in autonomously reasoning, generating task instructions, and allocating tasks. These challenges are further exacerbated by the need to address complex temporal, spatial, and heterogeneous task constraints. To address these limitations, inspired by the success of the Large Language Models (LLMs) in natural language understanding and logical inference, this paper proposes an environment-driven and LLM-guided task inference and allocation framework with dual-system temporal logics. The framework consists of three key modules: the Environment Module, which employs environment LTL to continuous monitor and verify environmental resource constraints; the Inference Module, leveraging LLMs for autonomous generation and verification of robotic tasks in response to resource changes; and the Robot Module, which explores the feasible task allocation for the multi-robot system. When the resources in the environment do not satisfy the specification, the Inference Module is used to analyze and infer the feasible actions to be executed by the multi-robot system, so as to alleviate the environmental resource problem. Experimental results demonstrate the scalability, efficiency, and autonomy of our framework across varying task environments and robot configurations. Note to Practitioners—This work presents an environment-driven and LLM-guided framework for multi-robot task inference and allocation, providing practical guidance for deploying autonomous decision-making in dynamic domains such as logistics, manufacturing, and agriculture. We recognize that successful real-world implementation requires deep integration with existing robotic middleware and systematic handling of three key challenges: seamless coordination among heterogeneous platforms, mitigation of real-time reasoning disturbances caused by intermittent communications, and robustness against perception noise that may affect temporal logic translation. In our design, heterogeneous capabilities are quantitatively modeled to facilitate robot cooperation, while a cyclic environment verification and “Generation–Verification–Regeneration” closed loop mechanism are introduced to alleviate communication delays and interference. Furthermore, the framework’s modular architecture, comprising continuous verification of environmental requirements, autonomous inference of task strategies, and optimized allocation among robot teams, ensures high flexibility and extensibility. This design allows practitioners to embed robustness enhancements (e.g., communication retry mechanisms, state estimation filters) and safety assurance modules according to specific deployment needs, without altering the core logical structure. Numerical and simulation studies have preliminarily validated the scalability of the framework in large-scale task settings. Beyond a theoretical construct, this work provides a transferable methodological foundation for building large-scale, environment-aware, and cooperatively intelligent multi-robot systems.
Keywords:
Linear temporal logic
large language models
multi-robot task allocation
task inference
Journal
IF:
6.4
Papers:
4.9K
Citations:
1.6W