Avionics News February 2023 - 37

Learning assurance process
aircraft there or not, and where is it exactly in the image. "
There are literally millions of parameters the algorithm
counts in these statistical equations, which is why no human
is able to monitor it and influence what it does. This is the
role of a training algorithm to compare its answers to the ones
provided by humans in annotated images until the neural
network has reliability similar to the humans.
" At this stage, we freeze the code, " according to
Daedalean, and the neural network doesn't change anything
after that on that version. The code is then released to partners
such as aircraft companies for evaluation. They give feedback
on what needs to be adjusted in the next release, and these
happen several times per year.
The Daedalean software development plan is to have the
AI develop alternative maneuvers to the established flight
plan to avoid hazards, then this information will be interfaced
with the flight controls. With full autonomy, this will be
used to automatically adapt the flight to avoid obstacles
and terrain. The pilot's role will become limited and then
there will be no pilot aboard at all, according to Ramanujam
Kalale, a certification specialist at Daedalean. Kalale has
a background in avionics and certification and worked at
Pilatus Aircraft for several years as an avionics and electrical
engineer.
" Eventually, the plan is to perform risk analysis to provide
probable action a pilot might take and then infuse this
information into the aircraft controls and also communicate
with air traffic control on the ground, " Kalale said. The
system will also communicate vehicle to vehicle within fleets
of trusted aircraft with a fully autonomous AI pilot capability
being available in five to seven years.
While machine learning is used to develop the system on
the ground, once it is in operational service, it will not evolve
on the flight deck. Operational flight data are used back on
the ground with machine learning to improve Daedalean's
algorithms for the next software release. Simulated data are
also used as it increases the amount of data available to train
the algorithms. The company's custom-designed simulator
takes landscapes from real flights and adds different weather
and hazards to the images that multiply them by an order of
magnitude for testing and training.
" We are by far the most advanced in getting to this level
of detail for design and certification of machine learning,
artificial intelligence systems, " Kalale said. " We have
competitors and work with some of them. Not all of them use
machine learning, and we are working with them to test our
technology in their products as well. Very few are working on
machine learning for these avionics purposes. "
Another one of the largest avionics companies in the
United States is also working with Daedalean. One AEA
member company, Honeywell, has signed an agreement on
joint testing and technological partnership with Daedalean
in developing solutions for autonomous takeoff, landing and
GPS-independent navigation plus collision avoidance for
general aviation aircraft and eVTOL air vehicles. In addition,
Honeywell Ventures has joined the Swiss startup's pool of
investors, but the amount and conditions of this arrangement
have not been disclosed. No further details on this partnership
are currently available.
In August 2022, Daedalean appointed Dr. Yemaya
Bordain as president of Daedalean Americas business based
in Phoenix, Arizona. She was previously an electrical and
computer engineer and global sales director, public sector,
Internet of Things Group at Intel. q
avionics news * february 2023 37
Avionics News February 2023

Table of Contents for the Digital Edition of Avionics News February 2023
