Deep reinforcement learning-based multi-agent negotiation framework for conflict-free U-Space scenarios

Very-Low Level urban airspace is expected to accommodate large numbers of simultaneous drone missions in the near future, posing acute challenges for the strategic planning and deconfliction services mandated by U-space regulations. The current strategic planner algorithm relies on the first-come-first-served or batch policy that rejects nearly half of the submitted flight plans in large-scale scenarios. This paper presents a novel study that integrates a Deep Reinforcement Learning-based Multi-Agent Negotiation Framework into the U-space strategic planning and deconfliction service. The framework reallocates rejected flight plans through an iterative sealed-bid auction mechanism in a fully competitive auction environment. Four role-specific drone operator agents were constructed: Achiever, Optimizer, Economizer and Normalizer, to learn bidding policies with the Proximal Policy optimization algorithm. Each round’s reward combines individual profit, bidding cost and a global utilisation signal, guiding agents towards system-efficient yet self-interested behaviour. The simulation experiments show that the Deep Reinforcement Learning-based Multi-Agent Negotiation Framework lifts the mission re-accommodation ratio to a notable level while converging to a final allocation within a short strategic planning time. Analysis of reward distribution and bidding patterns confirms that agents adapt their strategies to heterogeneous operational objectives without explicit coordination. These findings indicate that learning-enabled auctions can reconcile operator competition with network-level efficiency, offering a scalable path towards conflict-free, high-density U-space operations.