Page 30 - World Airnews Magazine April 2020 Edition

DRONES DRONES

from different angles. not specifically trained for this purpose, it seems the rigorous
ROBUST DRONES ARE of real-world training. Like training wheels, learning in simulation simulation training also equipped the controller to follow a moving

Over the course of 24 hours, the system processes 250 hours
target precisely.

mode allows the drone to learn on its own in a safe environment, The researchers note that there's still work to be done. In this ex-
TEACHING THEMSELVES before being released into the wild. Eventually, it finds solutions to periment, they manually adjusted a few parameters on the drones,

for example, limiting maximum thrust, but the next step is to make
every challenge put in its path.

"In simulation we can run hundreds of thousands of scenarios," the drones completely independent. The experiment is a promising
move towards building sturdy drones that can tune themselves and
HOW TO FLY said Molchanov. learn from experience.
Professor Sukhatme, Molchanov's advisor and a Fletcher Jones
A REAL-WORLD CHALLENGE Foundation Endowed Chair in Computer Science, said the research

To prove their approach, the researchers moved the trained solves two important problems in robotics: robustness and gener-

controller onto real drones developed in Ayanian's Automatic alisation.

Coordination of Teams Lab. In a netted indoor drone facility, they "From a safety perspective, robustness is super important. If

flew the drones and tried to throw them off by kicking and pushing you're building a flight control system, it can't be brittle and fall

them. apart when something goes wrong," said Sukhatme.

The drones were successful in correcting themselves from mod- "The other important thing is generalisation. Sometimes you
erate hits (including pushes, light kicks and colliding with an object) might build a very safe system, but it will be very specialised. This

90% of the time. Once trained on one machine, the controller was research shows what a mature and accomplished Ph.D. student can

able to quickly generalize to quad copters with different dimen- achieve, and I'm very proud of Artem and the team he assembled."
sions, weights and sizes. Q

While the researchers focused on robustness in this study, they Article courtesy: https://techxplore.com/news/2020-03-ad-

were surprised to find the system also performed competitively in vanced-series-robust-drones.html
terms of trajectory tracking - moving from point A to B to C. While
IATA
by Caitlin Dawson, University of Southern
California GLOBAL SUSPENSION
LEARNING TO FLY
Roboticists have been turning to birds for flight inspiration for OF SLOT RULES

rones, specifically quad copters, are an adaptable lot. years. But drones have a long way to go before they're as agile
DThey've been used to assess damage aft er disasters, as their feathered counterparts. When a drone ends up in an

undesirable orientation, such as upside down, it can be difficult for

deliver ropes and life-jackets in areas too dangerous for ground- it to right itself. • Future bookings are softening and services in order to maintain slots. Air-

craft can be reallocated to other routes

based rescuers, survey buildings on fire and deliver medical "A drone is an inherently unstable system," said Molchanov. carriers are reacting with measures or parked; crew can have certainty on

specimens. "Controlling a drone requires a lot of precision. Especially when he International Air Transport such as crew being given unpaid their schedules.

But to achieve their full potential, they have to be tough. In something sudden occurs, you need a fast and precise sequence of TAssociati on has contacted avi- leave, freezing of pay increases, and “IATA research has shown that traffic

the real world, drones are forced to navigate uncertain shapes control inputs." ation regulators worldwide to request plans for aircraft to be grounded. has collapsed on key Asian routes and
in collapsing buildings, avoid obstacles and deal with challenging that the rules governing use of airport Given these extraordinary circumstanc- that this is rippling throughout the
But, if a drone was able to learn from experience, like humans, it
conditions, including storms and earthquakes. would be more capable of overcoming these challenges. slots be suspended immediately and for es as a result of the public health emer- air transport network globally, even

At the University of Southern California Viterbi school of en- the 2020 season, due to the impact of gency, the collective view of the airline between countries without major

gineering's department of computer science, researchers have With this is mind, the USC researcher team created a system that COVID-19 (the Corona virus). industry is that the application the 80% outbreaks of COVID-19. There are prec-

created artificially intelligent drones that can quickly recover when uses a type of machine learning, a subset of artificial intelligence, Around 43% of all passengers depart rule during the upcoming season is inap- edents for previous suspension of the
called reinforcement learning to train the drone in a simulated en-
pushed, kicked or when colliding with an object. The autonomous vironment. Or more precisely, to train the drone's ‘brain’ or neural from over 200 slot co-ordinated airports propriate. Flexibility is needed for airlines slot use rules and we believe the circum-
drone "learns" how to recover from a slew of challenging situations network controller. worldwide. At present, the rules for slot to adjust their schedules according to stances again calls for a suspension to

thrown at it during a simulation process. allocation mean that airlines must oper- extraordinary demand developments. be granted.”

"Currently, the controllers designed to stabilize quad copters "Reinforcement learning is inspired by biology - it's very similar ate at least 80% of their allocated slots Regulators have already been waiving “We are calling for regulators
require careful tuning and even then, they are limited in terms of to how you might train a dog with a reward when it completes a under normal circumstances. the slot rules on a rolling basis during worldwide to help the industry plan
command," said Molchanov.

robustness to disruption and are model-specific," said the study's Failure to comply means the airline los- the COVID-19 crisis primarily for for today’s emergency, and the future

lead author Artem Molchanov, a Ph.D. in computer science candi- Of course, drones don't get snacks. But in the process of rein- es its right to the slot the next equivalent operations to China and Hong Kong recovery of the network, by suspending

date in USC's Robotic Systems Embedded Laboratory. forcement learning, they do receive an algorithmic reward: a math- season. In exceptional circumstances, SAR. However, given the recent further the slot use rules on a temporary basis,

"We're trying to eliminate this problem and present an approach ematical reinforcement signal, which is positive reinforcement that regulators can relax this requirement. outbreaks this is no longer contained to “said Alexandre de Juniac, IATA’s direc-

that leverages recent advancement in reinforcement learning so it uses to infer which actions are most desirable. The COVID-19 crisis has had a severe im- the Asia markets. tor general and CEO.
we can completely eliminate hand-tuning controllers and make LEARNING IN SIMULATION pact on air traffic. Airlines are experienc- Without certainty that these waivers “The world is facing a huge challenge

drones super robust to disruptions." The drone starts in simulation mode. At first, it knows nothing ing serious declines in demand including: will continue for the summer season (or to prevent the spread of COVID-19 while

winter season in the Southern hemi-

The paper, called "Sim-to-(Multi)-Real: Transfer of Low-Level about the world or what it is trying to achieve, said Molchanov. It • A carrier experiencing a 26% reduc- enabling the global economy to contin-

Robust Control Policies to Multiple Quad rotors," was presented at tries to jump a little bit or rotate on the ground. tion across their entire operation in sphere), airlines are unable to plan ahead ue functioning. Airlines are on the front

sufficiently to ensure efficient rostering

the International Conference on Intelligent Robots and Systems. Eventually, it learns to fly a little bit and receives the positive re- comparison to last year of crew or deployment of aircraft. line of that challenge and it’s essential

Co-authors were Tao Chen, USC computer science master's inforcement signal. Gradually, through this process, it understands • A hub carrier reporting bookings to that the regulatory community work

Suspending the requirement for the
student; Wolfgang Honig, a former USC computer science Ph.D. how to balance itself and ultimately fly. Then, things get more Italy down 108% as bookings collapse entire season (to October 2020) has with us to ensure airlines are able to

student; James A. Preiss, a computer science Ph.D. student; Nora complicated. to zero and refunds grow meant that airlines can respond to mar- operate in the most sustainable manner,
Ayanian, USC assistant professor of computer science and Andrew While still in simulation, the researchers throw randomized con- • Many carriers reporting 50% no- ket conditions with appropriate capacity both economically and environmentally,

and Erna Viterbi Early Career Chair; and Gaurav Sukhatme, profes- shows across several markets to alleviate the worst impacts of the

sor of computer science and electrical and computer engineering ditions at the controller until it learns to deal with them successful- levels, avoiding any need to run empty crisis,” he said. Q
ly. They add noise to the input to simulate a realistic sensor. They

and USC Viterbi executive vice dean.
change the size and strength of the motor and push the drone
World Airnews | April 2020 World Airnews | April 2020
— 28 — — 29 —