Stochastic Agent-Based Metaheuristics for Distributed Task Allocation in Heterogeneous Robotic Swarms with Partial Observability

Abstract

Swarm robotic systems comprised of many relatively simple robots are often deployed in domains where tasks appear in a spatially distributed and time varying manner. Such environments include environmental monitoring, search and rescue operations, and warehouse logistics where the cost of centralised planning is high and communication may be unreliable. In these settings, heterogeneity in robot sensing, mobility, and manipulation capabilities complicates task allocation because agents cannot be treated as interchangeable. At the same time, local sensing and limited-range communication imply that individual robots operate under partial observability, with only fragmentary and delayed information about the global task configuration and the states of peers. This combination of heterogeneity, decentralisation, and uncertainty motivates distributed decision mechanisms that are lightweight, robust to missing information, and able to adapt online to evolving workloads. This paper develops and studies stochastic agent-based metaheuristics for distributed task allocation in heterogeneous robotic swarms under partial observability. Each robot is modeled as an autonomous decision maker executing a probabilistic policy that balances local exploitative allocation decisions with exploratory behavior guided by metaheuristic principles. The proposed framework couples a linear assignment relaxation, used as an abstract global benchmark, with local learning rules that update task preferences using noisy observations and intermittent communication. Analytical arguments and extensive simulated scenarios are used to examine how the algorithmic parameters shape convergence speed, load balancing, and resilience to sensing limitations. Emphasis is placed on understanding trade-offs between exploration, communication density, and heterogeneity-aware coordination rules, rather than on demonstrating a single optimal design. The results illustrate characteristic behaviors of stochastic metaheuristics in partially observed swarm environments and highlight design considerations for future task allocation mechanisms in heterogeneous multi-robot systems.