A Unified Decentralized Framework for Task Allocation in Heterogeneous Multi-Robot Systems

Abstract

Effective allocation of tasks in heterogeneous multi-robot systems (MRS) is a pivotal challenge as real-world deployments increasingly feature diverse robotic agents with varied capabilities, resources, and mobility constraints. Achieving scalable, robust, and efficient collaboration in such environments is impeded by the combinatorial complexity of task allocation, the need for real-time adaptability, and the possibility of dynamic changes such as agent failures or shifting team compositions. While centralized approaches can solve small-scale problems, they falter in large teams due to high computational and communication overhead, lack of resilience, and suboptimal adaptability to environmental changes or agent faults. Recent work in coalition game theory, deep reinforcement learning (RL), and distributed architectures has made significant progress in these areas—offering decentralized, learning-driven methodologies for task allocation and scheduling. However, a fully unified framework that reliably addresses agent heterogeneity, asynchrony, coalition formation, and failure handling in a modular and generalizable way remains an open challenge.
In this work, we synthesize the state-of-the-art, drawing from coalition-game utility modeling, RL-based decentralized scheduling, asynchronous multi-agent RL, and ROS-enabled distributed architectures, to propose a unified decentralized framework for robust task allocation in heterogeneous multi-robot teams. Our architecture integrates decentralized policy engines, dynamic coalition negotiation, macro-action–based asynchrony, motivation-driven task reallocation, and communication middleware, allowing teams to self-organize, adapt, and recover from failures without central control. This design enables scalable deployment in realistic environments, supports online learning and adaptation to unforeseen tasks or conditions, and minimizes both resource contention and deadlocks through predictive commitment and masking mechanisms. Planned simulation and physical experiments will validate our approach on metrics such as global task completion, resource efficiency, makespan, and fault tolerance. Our unified framework not only harmonizes the major advances in multi-robot task allocation literature but also positions itself as a practical blueprint for real-world heterogeneous MRS deployments in domains such as disaster response, industrial automation, and exploration.