Abstract
Automating Air Traffic Control (ATC) requires capturing the nuanced, "unwritten rules" that human controllers use to manage complex airspace. This paper proposes an Inverse Optimal Planning framework that utilizes Maximum Entropy Inverse Reinforcement Learning (MaxEnt IRL) to extract cost functions from real-world flight data. By integrating these learned costs into a search-based motion planner (ARA*) using a Dubins Airplane model, the system generates trajectories that are both safe and human-like. Evaluation on arrival data from Seattle-Tacoma International Airport shows that the learned model reduces path imitation error from 4 km to 1 km and significantly improves planning efficiency.
1. Introduction
Air Traffic Control is a high-stakes domain where human operators must balance safety, efficiency, and complex airport-specific protocols. As airspace becomes increasingly crowded with the rise of urban air mobility, human-centered bottlenecks are becoming more critical. Traditional automated methods often rely on rigid, hard-coded rules that fail to capture the qualitative strategies used by experts. This research addresses the challenge of "Autonomous ATC" by learning these strategies directly from historical flight data, enabling a planner to mimic expert decision-making.
2. Methodology
The proposed framework, Inverse Optimal Planning, consists of three core components:
Vehicle Modeling: The system employs the Dubins Airplane model, a 4D state space $(x, y, z, \text{bearing})$ that assumes a fixed speed. This provides a computationally efficient yet realistic approximation of commercial aircraft dynamics.
Motion Planning: Pathfinding is performed using Anytime Repairing A (ARA). To handle the specific requirements of aviation, a custom heuristic was developed to account for helical paths necessary for altitude reduction.
Inverse Reinforcement Learning: The core of the approach is the use of Maximum Entropy IRL to recover the underlying cost function $J$ from expert demonstrations. The cost function is split into two primary components:
- $J_a$ (Airport Routing): A spatial potential field, stored in a sparse hash-map, that learns preferred arrival corridors.
- $J_o$ (Inter-Airplane Safety): A cylindrical penalty function that enforces horizontal and vertical separation.
3. Results
The model was evaluated using FlightAware data for arrivals at Seattle-Tacoma International Airport (SEA).
3.1 Imitation Accuracy
The planner using the learned cost function significantly outperformed a standard "shortest path" baseline. As shown in Figure 1, the average minimum path difference between the generated trajectories and actual human-controlled flights was reduced from approximately 4 km to 1 km.
3.2 Qualitative Analysis
Qualitative results demonstrate that the learned cost function $J_a$ guides the planner toward standard arrival routes, avoiding the unnatural maneuvers produced by simple path-length minimization.
3.3 Computational Efficiency
The learning process also improved the search efficiency of the ARA* planner. At the beginning of training, the planner timed out on nearly 100% of tasks due to an uninformative cost landscape. After 8,000 training steps, the time-out rate dropped to roughly 30%, indicating that the learned cost function effectively guides the heuristic search.
4. Conclusion
This work demonstrates that Inverse Optimal Planning can successfully capture the complex preferences of human air traffic controllers. By decoupling the learning of "rules of the road" from the underlying vehicle dynamics, the framework offers a flexible path toward autonomous airspace management. Future research will focus on incorporating dynamic velocity models and exploring distributed control for multi-agent coordination.
Summary updated