Preliminary Design of Human-like Decentralised Task Assignment for Heterogeneous Unmanned Vehicles using Reinforcement Learning

Efficient coordination among unmanned aerial systems is essential in applications such as surveillance, search and rescue, environmental monitoring, and logistics, where tasks are distributed, time-critical, and executed under limited communication and perception. Ensuring robust and scalable task allocation in these conditions remains a major challenge. Recent advances in artificial intelligence, particularly reinforcement learning, provide promising tools to address it. Unlike classical optimisation or auction-based strategies, these approaches can encode cooperative behaviours through experience, enabling generalisable policies without relying on predefined coordination rules. This work presents a preliminary study on decentralised task assignment for teams of unmanned agents operating in dynamic environments, where agents perform a Human-like communicationdecision-agreement process. Each agent selects its task based on self-estimated execution cost, task-specific efficiency, and limited information exchanged with teammates. The latter property supports cooperation between agents, that is allowed but not strictly imposed in order to speed up tasks completion. The training framework is intentionally scenario-agnostic, incorporating stochastic cost variations and heterogeneous agent capabilities to foster generalisation. Early results demonstrate that the learned policy can yield coherent and cooperative allocations, and that human-like communication among agents promotes intuitively interpretable and stable task distributions. These findings highlight the potential of reinforcement learning to support scalable and resilient autonomy in multi-agent systems, paving the way for future validation in real-world scenarios.