Design-for-Failure in Multi-Robot Systems through Integrated Architectures and Proactive Successor Allocation

Abstract

Modern warehouse multi-robot systems (MRS) are required to sustain high operational efficiency while remaining resilient to inevitable robot failures. However, traditional multirobot task allocation (MRTA) research typically assumes fault-free environments. This assumption often results in centralized recovery bottlenecks or excessive computational overhead when failures occur. To address these limitations, this thesis introduces a novel binary analysis framework that categorizes fault-tolerant architectures into two types: native and integrated. Native architectures embed resilience directly into the task allocation design. In contrast, integrated architectures achieve fault tolerance through modular coupling with recovery mechanisms.

A high-fidelity 2D simulation platform was developed to perform benchmark evaluations of four representative algorithms: the centralized native algorithm MRPF, the distributed native algorithm BFTC, and two integrated methods: FT-CBPA and FT-ACO + BA. Experimental results under different load conditions demonstrated that integrated fault-tolerant architectures, which combine high-performance allocation strategies with modular recovery mechanisms, consistently outperform native architectures that tightly embed recovery logic within the allocation process. Comparative simulations further reveal that integrated architectures achieve superior scalability and higher task completion rates under failure scenarios.

Building upon the advantages identified in the integrated approach, this thesis proposes SPA-CBPA (Successor Pre-Allocation Consensus-Based Payload Allocation). This method transforms fault recovery from a reactive, post-event reallocation process into a deterministic pre-event activation mechanism through proactive successor assignment. Consequently, it successfully reduces the computational complexity of fault recovery from O(N) to almost O(1). Extensive benchmarking indicates that, compared to reactive benchmarks, SPACBPA reduces the delay in fault recovery to nearly zero and decreases the overhead of recovery-related communication by up to 65.2%. Although instantaneous recovery results in a modest 20% to 38% increase in total makespan, the proposed method significantly improves predictability and operational reliability in mission-critical logistics environments.

Overall, this research concludes a “design-for-failure" paradigm, demonstrating that proactive successor pre-allocation is essential for developing the next generation of scalable, efficient, and resilient multi-robot systems.