Augmenting Flight Training with AI to Efﬁciently Train Pilots Michael Guevarra1 Srijita Das23 Christabel Wayllace23 Carrie Demmans Epp2 Matthew E. Taylor23 Alan Tay1

2025-04-24 0 0 1.32MB 3 页 10玖币

侵权投诉

Augmenting Flight Training with AI to Efﬁciently Train Pilots

Michael Guevarra1, Srijita Das2,3, Christabel Wayllace2,3, Carrie Demmans Epp2,

Matthew E. Taylor2,3, Alan Tay1

1Delphi Technology Corp: guevarrm@myumanitoba.ca,alantay@delphitechcorp.com

2University of Alberta: {srijita1, wayllace, demmanse, matthew.e.taylor}@ualberta.ca

3Alberta Machine Intelligence Institute (Amii)

Abstract

We propose an AI-based pilot trainer to help students learn

how to ﬂy aircraft. First, an AI agent uses behavioral cloning

to learn ﬂying maneuvers from qualiﬁed ﬂight instructors.

Later, the system uses the agent’s decisions to detect errors

made by students and provide feedback to help students cor-

rect their errors. This paper presents an instantiation of the

pilot trainer. We focus on teaching straight and level ﬂying

maneuvers by automatically providing formative feedback to

the human student.

There is a critical shortage of commercial pilots world-

wide: according to Oliver Wyman (2022), there will be a

global gap of 34,000 pilots by 2025. Part of the problem

is that pilots qualiﬁed to conduct such training are in very

high demand and in short supply. Currently, human instruc-

tors guide trainees using ﬂight simulator exercises. We posit

that training an AI-enabled system to provide instruction for

some tasks is a viable approach to reducing instructor work-

load while allowing them to interact with more students.

This could increase the number of students per pilot trainer,

improving the throughput of training pilots and therefore in-

crease the supply of trained pilots.

In recent human-in-the-loop research, AI agents use ad-

vice from humans in different forms to speed up learn-

ing (Bignold et al. 2021; Cui et al. 2021; Christiano et al.

2017; Da Silva*et al. 2020). Speciﬁcally, imitation learn-

ing allows an agent to learn to mimic a human’s behavior.

Further, AI has been used for airplane ﬂying (Morales and

Sammut 2004; Sandstr¨

om, Luotsinen, and Oskarsson 2022)

as well as inside intelligent tutoring systems for multiple

tasks ranging from student skill development (Georgila et al.

2019) to improving teaching strategies (Wang 2018). This

paper presents a system where an agent mimics a qualiﬁed

pilot and assists students in a pilot training program. Specif-

ically, we focus on the straight and level ﬂight task as a

proof of concept. A trained agent identiﬁes mistakes or sub-

optimal maneuvers of trainee pilots inside a ﬂight simulator

and suggests corrective actions. To the best of our knowl-

edge, this is a ﬁrst attempt to use an AI tutor to train human

pilots for ﬂight.

Straight and Level

Pilot BCAgent

Human Expert AI Teacher

Advice

Demonstrations

Feedback

X-Plane

Adaptive

Visual Feedback

Task

Flight Simulator

Students

Trainee Pilots

Figure 1: System Architecture

System Architecture

Our proposed intelligent tutoring system is shown in Fig-

ure 1. It includes four general components: (1) Task, (2)

Human expert, (3) AI teacher, and (4) Students. The task

is sampled from a curriculum of ﬂying maneuvers useful for

learning to ﬂy. Both the pilot and students perform tasks us-

ing the ﬂight simulator X-Plane.1We use the fundamental

ﬂight maneuver “straight and level”2as the target task. There

is a human expert (pilot) who is adept in the task and can

provide advice to train the AI teacher. After the AI teacher

is trained to mimic the human teacher, the AI teacher is used

to guide students by providing different types of feedback

based on their performance on the target task.

Modeling the Pilot: Agent Training

We trained a decision-making agent to learn from

pilot demonstrated trajectories. Experts demonstrated

the“straight and level task” inside the ﬂight simulator for

12.5minutes. They operate under visual ﬂight rules, i.e,

clear weather to ﬂy towards the target. The target direction is

changed between every trial demonstrated by the pilot to ac-

count for diversity in trajectory collection. The ﬁnal dataset

consists of 25 trials of 30 seconds each with a randomized

1https://developer.x-plane.com/article/airport-data-apt-dat-ﬁle-

format-speciﬁcation/

2Straight and level ﬂight is a ﬂight in which a constant head-

ing and altitude are maintained. It is an essential ﬂight maneuver

used to form correct habits. All other ﬂight maneuvers derive from

straight and level.

arXiv:2210.06683v1 [cs.LG] 13 Oct 2022

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

AugmentingFlightTrainingwithAItoEfcientlyTrainPilotsMichaelGuevarra1,SrijitaDas2,3,ChristabelWayllace2,3,CarrieDemmansEpp2,MatthewE.Taylor2,3,AlanTay11DelphiTechnologyCorp:guevarrm@myumanitoba.ca,alantay@delphitechcorp.com2UniversityofAlberta:fsrijita1,wayllace,demmanse,matthew.e.taylorg@ualberta.c...

展开>> 收起<<

Augmenting Flight Training with AI to Efﬁciently Train Pilots Michael Guevarra1 Srijita Das23 Christabel Wayllace23 Carrie Demmans Epp2 Matthew E. Taylor23 Alan Tay1.pdf

共3页,预览1页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

Augmenting Flight Training with AI to Efﬁciently Train Pilots Michael Guevarra1 Srijita Das23 Christabel Wayllace23 Carrie Demmans Epp2 Matthew E. Taylor23 Alan Tay1

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: