[Paper Review] Cooperative Decentralized Multi-agent Control under Local LTL Tasks and Connectivity Constraints

Meng Guo, Jana Tůmová+1 more|arXiv (Cornell University)|May 8, 2014

Formal Methods in Verification2 references4 citations

TL;DR

This paper proposes a decentralized control framework for multi-agent systems with local Linear Temporal Logic (LTL) tasks and connectivity constraints, using a dynamic leader-follower scheme to ensure task satisfaction and real-time connectivity. By avoiding synchronized product construction and enabling real-time leader switching, the method guarantees all agents achieve their individual and collaborative LTL specifications while maintaining communication links.

ABSTRACT

We propose a framework for the decentralized control of a team of agents that are assigned local tasks expressed as Linear Temporal Logic (LTL) formulas. Each local LTL task specification captures both the requirements on the respective agent's behavior and the requests for the other agents' collaborations needed to accomplish the task. Furthermore, the agents are subject to communication constraints. The presented solution follows the automata-theoretic approach to LTL model checking, however, it avoids the computationally demanding construction of synchronized product system between the agents. We suggest a decentralized coordination among the agents through a dynamic leader-follower scheme, to guarantee the low-level connectivity maintenance at all times and a progress towards the satisfaction of the leader's task. By a systematic leader switching, we ensure that each agent's task will be accomplished.

Motivation & Objective

• To address the challenge of decentralized control in multi-agent systems where each agent has a complex local LTL task requiring collaboration with others.
• To ensure all agents satisfy their individual LTL specifications, including collaborative behaviors, while maintaining communication connectivity at all times.
• To avoid the computationally expensive construction of synchronized product systems between agents, enabling scalable and distributed planning.
• To develop a real-time, distributed control scheme that integrates leader election with motion control for dynamic task execution.
• To guarantee long-term progress toward all agents' tasks through systematic leader re-election and decentralized coordination.

Proposed method

• Each agent generates a local high-level plan based on its individual LTL specification using automata-theoretic synthesis.
• A dynamic leader-follower coordination scheme is employed, where the current leader guides the team toward its task while maintaining connectivity.
• A distributed leader election algorithm ensures sequential progress toward each agent’s task using only local communication.
• A continuous decentralized controller is designed for the leader and followers, ensuring motion convergence and connectivity maintenance within a radius of 1.5m.
• Leader switching is systematically triggered upon task completion or milestone achievement, ensuring all agents eventually satisfy their tasks.
• The approach avoids global synchronization and product automaton construction, enabling real-time execution without offline planning.

Experimental results

Research questions

• RQ1 How can decentralized multi-agent control be achieved when each agent has a local LTL task that may require collaboration with others?
• RQ2 How can connectivity constraints be maintained in a decentralized multi-agent system during dynamic task execution?
• RQ3 Can a leader-follower scheme be designed to guarantee progress toward all agents’ LTL tasks without constructing a global product automaton?
• RQ4 What is a scalable, real-time coordination mechanism that ensures all agents eventually satisfy their individual and collaborative LTL specifications?
• RQ5 How can leader re-election be systematically managed to ensure long-term task satisfaction under communication constraints?

Key findings

• The system successfully maintains connectivity throughout the simulation, with all pairwise distances between neighboring agents staying below the 1.5m communication radius.
• Agent $τ_1$ reached its first goal region $R_{11}$ by $t=3.8s$, and all agents converged to $R_{21}$ by $t=6.1s$ after collaboration.
• The leader switched systematically through all agents: $τ_1$, $τ_2$, $τ_3$, $τ_4$, $τ_2$, $τ_3$, $τ_1$, $τ_1$, $τ_3$, $τ_3$, $τ_4$, $τ_4$, covering all required goal regions over 35 seconds.
• All agents satisfied their respective LTL tasks: $τ_1$ completed object H loading, $τ_2$ and $τ_3$ performed periodic surveillance, and $τ_4$ achieved repeated assembly of object C.
• The trajectory and distance plots confirmed that the decentralized controller maintained stable connectivity and achieved all task milestones in real time.
• The method successfully avoided the construction of a synchronized product system, enabling scalable and distributed execution without global synchronization.