Drone / eVTOL deconfliction

For time-critical medical deliveries, waiting for global airspace replanning isn’t an option. But neither is disrupting scheduled eVTOL traffic. A new learning-based approach shows how emergency UAVs can safely navigate dense terminal airspace, in part by learning to give way, writes Peter Donaldson.

The promise of advanced air mobility hinges on integrating diverse aircraft types into shared low-altitude airspace. But, as vertiport traffic intensifies, a critical operational tension emerges around the problem of how to extract a time-critical emergency UAV safely from congested terminal airspace without disrupting scheduled eVTOL traffic.

Researchers from Tsinghua University in Beijing have tackled precisely this question. Their approach frames the emergency egress problem as a hierarchical strategic–tactical conflict management challenge. Scheduled eVTOL operations – subject to strategic planning through air traffic management (ATM) – are treated as high-priority, non-cooperative traffic with fixed trajectories. The emergency UAV, by contrast, must find its own way out using limited onboard sensing and real-time decision making. The strategic schedule remains untouched; the tactical layer adapts.

The proposed solution employs Asynchronous Proximal Policy Optimisation (APPO), a distributed reinforcement learning algorithm implemented via Ray RLlib, an open-source, industry grade reinforcement learning library. Training takes place in a realistic simulation environment incorporating AirSim’s AirLib flight-control model, ensuring that the learned policy respects genuine UAV dynamics during execution of speed and heading commands.

The results reveal a counter-intuitive but highly effective learned strategy whereby the UAV frequently chooses to reduce speed when evading an oncoming eVTOL aircraft, rather than accelerating out of the way. This behaviour reflects the operational hierarchy – by yielding, the UAV allows the higher-priority, passenger-carrying eVTOL aircraft to maintain its predictable path. Reducing speed also tightens the UAV’s turn radius and increases the time-to-contact, giving the policy additional seconds to execute sequential adjustments in dense, fast-changing environments.

The APPO-based planner was benchmarked against five non-learning baselines: artificial potential fields, velocity obstacles, rapidly-exploring random trees, modified closed-loop rapidly-exploring random trees and Monte Carlo tree search with upper confidence bounds applied to trees. Under locally dense terminal traffic, the learning-based approach demonstrated greater robustness in safe-egress performance, separation maintenance and, crucially, online computational efficiency. It achieved substantially lower ‘near mid-air collision’ rates while maintaining corridor alignment and target-directed exit behaviour, the researchers report.

As reinforcement learning continues to prove its viability for real-time, safety-critical applications in aerospace, the challenge now lies in verification aligned with aviation software standards. But for the immediate challenge of safely extracting an emergency drone from a crowded terminal, this research shows that sometimes the smartest move is to slow down, let bigger aircraft pass by and then resume the mission.

Reference: Su, L., Chen, S., Wang, Y. et al. (2026). “Learning-based emergency egress trajectory planning for UAV in mixed terminal airspace.” Transportation Research Part E: Logistics and Transportation Review, 214, 105020.

UPCOMING EVENTS