Quantify the burden
Metrics capture how hard the avoidance system must work and how strongly it disrupts the airspace.
ATM 2019 · Airspace safety
Collision avoidance in dense airspace.
A deep reinforcement learning correction layer helps traditional collision-avoidance logic remain safe and efficient as traffic density rises.
Learned correction layer
04
DomainUnmanned airspace
MethodDeep RL correction
ObjectiveSafety + efficiency
VenueATM Seminar 2019
01 The problem
Safety logic meets traffic density.
New unmanned operations can push free-flight airspace toward traffic densities that traditional collision-avoidance systems were not designed to handle. More encounters create more alerts, more maneuvering, and a growing decision burden.
Instead of replacing an established avoidance system, this work learns a correction around it. Deep reinforcement learning adapts the system’s decisions for dense environments while preserving the safety foundation of the original logic.
Q / 04
How can proven collision-avoidance logic adapt when dense traffic breaks its operating assumptions?
02 The correction strategy
Preserve the base. Learn the adjustment.
The approach decomposes the global state into pairwise encounters, combines their avoidance values, and adds a learned correction. New metrics quantify both the decision burden placed on the avoidance system and its operational impact on the surrounding airspace.
Learn around proven logic
A deep reinforcement learning function adjusts the aggregated avoidance values for dense encounters.
Balance safety and flow
The corrected system is assessed for both high safety levels and efficient operation under dense traffic.
03 What it showed
Adaptation without discarding the safety core.
Learning can improve how a conventional system behaves at the edge of its design envelope.
High safety levels maintained
The corrected collision-avoidance system preserved high levels of safety in dense airspace.
Less operational disruption
Compared with traditional methods, the corrected approach operated more efficiently under high traffic density.
A measurable system tradeoff
Decision-burden and airspace-impact metrics make the balance between intervention and efficiency explicit.
04 / Paper
ATM Research & Development Seminar · 2019
Optimizing Collision Avoidance in Dense Airspace Using Deep Reinforcement Learning
Air Traffic Management Research and Development Seminar · 2019