Periodic Reporting for period 1 - ORCI (Optimised Runway Centreline Interception)
Reporting period: 2024-06-01 to 2025-05-31
Context
Air traffic controllers help planes land at increasingly busy airports. They need to keep planes as close together as safely possible to maintain capacity. The EU-funded ORCI project aims to use AI solutions to improve runway throughput and efficiency, especially during periods of medium to very high traffic density. The project plans to develop an advanced automation tool for Terminal Maneuvering Areas. This tool will be based on an AI model trained on surveillance data and other data derived from air traffic control voice communications between pilots and controllers.
The project will ultimately demonstrate and validate this solution, potentially revolutionizing efficiency and performance in air traffic management.
Overall Objectives
Developing a Machine Learning model to support ATCo’s tactical decisions when vectoring aircraft to intercept the final approach localizer in the most optimal manner.
Reducing ATCo workload by minimizing the number of vectoring instructions in the final approach, as well as supporting mental processes focusing on aircraft turns and spacing estimations.
Improving runway throughput by delivering optimized arrival control services customized to specific approach procedures and airspace user behaviors.
Air traffic controllers help planes land at increasingly busy airports. They need to keep planes as close together as safely possible to maintain capacity. The EU-funded ORCI project aims to use AI solutions to improve runway throughput and efficiency, especially during periods of medium to very high traffic density. The project plans to develop an advanced automation tool for Terminal Maneuvering Areas. This tool will be based on an AI model trained on surveillance data and other data derived from air traffic control voice communications between pilots and controllers.
The project will ultimately demonstrate and validate this solution, potentially revolutionizing efficiency and performance in air traffic management.
Overall Objectives
Developing a Machine Learning model to support ATCo’s tactical decisions when vectoring aircraft to intercept the final approach localizer in the most optimal manner.
Reducing ATCo workload by minimizing the number of vectoring instructions in the final approach, as well as supporting mental processes focusing on aircraft turns and spacing estimations.
Improving runway throughput by delivering optimized arrival control services customized to specific approach procedures and airspace user behaviors.
During the initial reporting period, the project advanced significantly towards its objectives of enhancing arrival management through AI integration and human-machine interaction. Key activities focused on understanding operational environments, processing surveillance data, developing AI-based predictive models, and establishing a robust technical architecture to support future validation.
A detailed assessment was conducted on the operational characteristics of the final approach procedures at Barcelona (LEBL) and Lisbon (LPPT) airports. This analysis supported the development of the initial Operational Services and Environment Description (OSED).
Historical surveillance data was sourced from ENAIRE (LEBL) and NAV-PT (LPPT), covering three months of radar data for each airport. Relevant data segments were extracted, cleaned, and categorized for AI training and testing purposes. For LEBL, the focus was on Trombone operations to runway 24R in segregated mode (arrival only), while for LPPT, attention was given to Point Merge operations with runway 02 active in mixed mode. Datasets were split into 75% training and 25% testing subsets.
Using the prepared datasets, various machine learning techniques—including Linear Regression, Random Forest, and XGBoost—were explored. Linear Regression emerged as the most promising model, offering a balance between predictive accuracy and interpretability, aligning with the project's goal of Explainable AI (XAI).
For Barcelona (LEBL), two distinct models were developed:
Model 1: Predicting aircraft behavior during turn initiation (dynamic phase)
Model 2: Predicting behavior as aircraft stabilize on the localizer (steady phase)
For Lisbon (LPPT), two models were also developed, triggered based on the follower aircraft's assigned Point Merge Segment (Eastern vs. Western legs). Unlike Barcelona, the models for Lisbon operate in parallel, not sequentially.
To facilitate hybrid validation (Human-in-the-Loop and model-based), a service-oriented architecture was designed and implemented, enabling interoperability among all ORCI components through a publish-subscribe messaging system hosted on AWS. This architecture supports real-time data exchange and component orchestration across cloud infrastructure.
The RAMS Plus simulation tool was upgraded to automatically detect and process relevant aircraft pairs for analysis. Successful initial integration tests between the AI components and RAMS Plus for LEBL were completed in May 2025, with full validation testing planned for the next reporting period (Barcelona: June/July; Lisbon: Sept/Oct).
A detailed assessment was conducted on the operational characteristics of the final approach procedures at Barcelona (LEBL) and Lisbon (LPPT) airports. This analysis supported the development of the initial Operational Services and Environment Description (OSED).
Historical surveillance data was sourced from ENAIRE (LEBL) and NAV-PT (LPPT), covering three months of radar data for each airport. Relevant data segments were extracted, cleaned, and categorized for AI training and testing purposes. For LEBL, the focus was on Trombone operations to runway 24R in segregated mode (arrival only), while for LPPT, attention was given to Point Merge operations with runway 02 active in mixed mode. Datasets were split into 75% training and 25% testing subsets.
Using the prepared datasets, various machine learning techniques—including Linear Regression, Random Forest, and XGBoost—were explored. Linear Regression emerged as the most promising model, offering a balance between predictive accuracy and interpretability, aligning with the project's goal of Explainable AI (XAI).
For Barcelona (LEBL), two distinct models were developed:
Model 1: Predicting aircraft behavior during turn initiation (dynamic phase)
Model 2: Predicting behavior as aircraft stabilize on the localizer (steady phase)
For Lisbon (LPPT), two models were also developed, triggered based on the follower aircraft's assigned Point Merge Segment (Eastern vs. Western legs). Unlike Barcelona, the models for Lisbon operate in parallel, not sequentially.
To facilitate hybrid validation (Human-in-the-Loop and model-based), a service-oriented architecture was designed and implemented, enabling interoperability among all ORCI components through a publish-subscribe messaging system hosted on AWS. This architecture supports real-time data exchange and component orchestration across cloud infrastructure.
The RAMS Plus simulation tool was upgraded to automatically detect and process relevant aircraft pairs for analysis. Successful initial integration tests between the AI components and RAMS Plus for LEBL were completed in May 2025, with full validation testing planned for the next reporting period (Barcelona: June/July; Lisbon: Sept/Oct).
Several foundational activities were initiated and progressed during the first reporting period. These efforts focused on the definition of operational scenarios, the design of the experimental framework, and early development of the interoperable validation platform architecture.
Results are expected in early 2026 when the full validation exercises involving operational ATCOs will be archieved.
Results are expected in early 2026 when the full validation exercises involving operational ATCOs will be archieved.