Airspace Encounter Models Roadmap 2023

The new 2023 Airspace Encounter Models roadmap was designed to support integration of drones and advanced air mobility (AAM) aircraft into the airspace. This directly accomplishes goals of the Federal Aviation Administration (FAA) Aircraft Certification Service (AIR) and UAS Integration Office (AUS). To support industry, the roadmap is aligned to support RTCA, Inc. SC-147 and SC-228 activities to develop detect and avoid (DAA) and collision avoidance systems to mitigate collision risk in Class B and the airspaces surrounding the busiest airports. The majority of development by MIT LL has been tasked by the FAA UAS Integration Office. Specifically, the FAA A11L.UAS.2 research task is the primary sponsor of development by MIT Lincoln Laboratory (MIT LL) and they have been the primary sponsor since 2019.

You can find a link to the most current roadmap on the masthead at the top of the Airspace Encounter Models website (Link) or by using this?link. A blog post explaining the current technical direction and a review of what was achieved under the previous roadmap can be found using this link. This LinkedIn post copies elements from the roadmap blog post.

Technical Objectives for 2023 Roadmap

The next 12-18 month roadmap for airspace encounter model development was strongly influenced by the lessons learned from training the initial terminal model described in?Correlated Bayesian Model of Aircraft Encounters in the Terminal Area Given a Straight Takeoff or Landing?and three recent articles:?Inferring Traffic Models in Terminal Airspace from Flight Tracks and Procedures?by Soyeon Jung and Mykel Kochenderfer ;?Deep generative modelling of aircraft trajectories in terminal maneuvering areas?by Timothé Krauth et al.; and?Leveraging Public Aeronautical Data to Characterize Aircraft Traffic Intent?by Andrew Weinert et al.

As part of the FAA A11L.UAS.2 research task, the 2023 roadmap has been aligned to support RTCA SC-147 and SC-228 activities to develop DAA and collision avoidance systems to mitigate collision risk in Class B airspaces surrounding the busiest airports, such as Los Angeles International Airport (LAX), or in the low altitude airspace of metropolises, like Los Angeles, Miami, or New York City. Because Class B airspace is surrounded by a Mode C veil (with requirements for transponders and ADS-B Out equipment), where many low altitude drone operations are planned, we anticipate 2023 roadmap activities will also support the ASTM International F38 activities, and specifically revisions of the ASTM DAA performance standard (ASTM F3442/F3442M) and planned first publication of test methods for DAA (ASTM WK62669).

Technical Approach

Our roadmap was designed to facilitate contributions of a trained Dynamic Bayesian Networks to model correlation between aircraft while further supporting the broader research and safety community by precalculating airspace-based metadata and human annotated datasets. Directed by the FAA UAS Integration Office, these contributions are first intended to directly support RTCA SC-147 and RTCA SC-228 and then by the broader community response to the expected FAA notice of proposed rulemaking (NPRM) for drone regulations.

Our technical approach will continue to prioritize generalizability over location specific models. This was informed by prior research that the distribution of observed aircraft speed, vertical rate, and turn rate, near two different final approach fixes (FAF) for different Class B airports, were similar given distance from the FAF and altitudes prescribed by instrument approach procedures (IAPs).?Slide 35 from the research presentation at ICNS 2022?illustrates that speed distribution, given an altitude range based on the prescribed altitude from the FAFs of?MILTT for BOS?and?COWWE for EWR, are similar.

We hypothesized by supplemental 4-D (time, latitude, longitude, altitude) track positions with airspace-based features, such as distance away from runway or distance from a FAF, generalized clusters of behaviors could be identified. Accordingly, the ICNS 2022 presented research was supported by a training dataset of machine annotated tracks. For each track point, we calculated the distance from waypoints, FAA reported single latitude and latitude for airports, the distance from end of runways, the relative angular difference between the true heading of the track and runway heading, and the difference between the track’s altitude and a runway’s visual glide slope. While these annotations are straightforward, to calculate these annotations for billions of track points for multiple airports over multiple months required a parallelized workflow using high performance supercomputing resources, like the Lincoln Laboratory Supercomputing Center (LLSC); a capability available to few organizations.

Our approach is threefold then. First is to train a Dynamic Bayesian Network to model the geometry of two or more aircraft at the start, conclusion, and closest point of approach of a correlated encounter. This model can then be used to seed correlated trajectories of multiple aircraft from any type of model. A separate Dynamic Bayesian Network model could be used to generate trajectories but conceptually, trajectories could be sampled from other models, such as GMMs or VAEs, based on the seed sampled from the correlated geometry model. We believe this approach offers the valued understandability of Dynamic Bayesian Networks while enabling encounter modeling to leverage recent AI/ML advances.

Second is to release the dataset used to train this model, along with supplemental metadata and machine annotated features. Based on the research presented at ICNS 2022, we will machine annotate the relative geometry between a track position and all nearby airspace features. This provides us with a robust feature space to train the correlated geometry model but not all calculated features will ultimately be used for the new model. By releasing all features to the community, we will reduce the risk for the community to build upon our research and explore how these additional features can improve encounter modeling in general. Lastly, we are planning to release a human annotated dataset to enable high confidence supervised learning.

Lastly, we will publish guidelines and a checklist on how to leverage the encounter models based on best practices established for RTCA and the Department of Defense. This documentation is intended to help regulators review the use of the models as part of safety cases and waiver applications and also to educate users on how to successfully and appropriately sample the models to create encounters.

Highlights Since Publishing Previous Roadmap

To summarize developments, since publishing the last roadmap, we highlight four peer-reviewed papers, two primarily authored by MIT LL encounter model team and two authored by others; and one technical report. These articles all cite previous encounter model literature. This list is not comprehensive and there is other literature that have recently leveraged the encounter models.

For each paper we highlight encounter model related technical contributions and how the research informed policy and standardization efforts, particularly to support objectives of the FAA AUS UAS Research, Engineering, & Analysis Division (AUS-300) and FAA AUS Safety and Integration Division (AUS-400). Furthermore applicants submitting waivers and safety cases to the FAA Aircraft Certification Service and Flight Standards Service are leveraging encounter model investments made by the FAA UAS Integration Office.

The discussion articles are:

• Correlated Bayesian Model of Aircraft Encounters in the Terminal Area Given a Straight Takeoff or Landing by Andrew Weinert , Ngaire Underhill , Christine (Chen) Serres , Randal Guendel
• Near Midair Collision Analog for Drones Based on Unmitigated Collision Risk by Andrew Weinert , Luis Alvarez , Michael Owen , Benjamin Zintak
• Collision Risk and Operational Impact of Speed Change Advisories as Aircraft Collision Avoidance Maneuvers by Sydney Katz , Luis Alvarez , Michael Owen , Samuel Wu , Marc Brittain , Anshuman Das , Mykel Kochenderfer
• Identifying Multithreat Encounters for Collision Avoidance Algorithm Validation Using Multiattribute Utility Theory by Yang Hu and Xiaoyan Wang
• sUAS Mid Air Collision Likelihood - Final Report by Armando De Abreu , Gerardo Arboleda , Gerardo Olivares , Luis Gomez , Deepak Singh , Timothy Bruner , Tom Haritos, Ph.D. , Katherine (Silas) Ragnoli , Ryan Wallace, Ed.D. , Andrew Weinert