Teaching a computer to refuel an aeroplane

September 13, 2019

Flying an aeroplane is
difficult enough. And one of the most difficult
operations a pilot can train for is air-to-air refuelling. But what we’re trying to do is
teach a computer to do it. Air-to-air refuelling is
essentially formation flying at close quarters. The receiver sidles
up to the tanker. And he tries to get
his probe into the basket, into the basket. It takes a lot of skill
to do this. And pilots train on a weekly
basis, just to keep their skills up. And when it goes wrong,
boy, does it go wrong. [CRASHING] To automate air-to-air
refuelling, there are two key technologies that need
to be developed. The first is sensors to
detect the position of the end of the hose. The second is control
algorithms, to guide the receiver aircraft to
that position. And this is the kit we’re going
to be using to do that. These are standard robotic arms,
the sort of thing you’d find in a car production line. But they’ve got 6 degrees of
freedom, which means you can manipulate something in
three dimensional space, in any position. So you’ve got three degrees of
freedom like this and then any orientation, by rotating
about different axes. Now, this is accomplished
using six joints. The first joint is at the base
here, which allows the whole robot to swivel around there. The second joint is here. The third joint is
up at the elbow. The fourth joint is about the
axis of the forearm here. The fifth joint is the
wrist, rotating here. And then, the sixth joint is
around the end effector here. Now, there’s another joint,
another axis, which is the track. And that gives us the room to
move backwards and forwards, extends the range
of the robots. And when they want to,
they can shift. So this is the refueling
probe. It’s mounted rigidly to
the receiver aircraft. The pilot has to manoeuvre this
probe into the reception coupling on the drove
assembly. And that’s when the
magic happens. So the basket hangs off the
end of a hose which is trailing from the
tanker aircraft. This canopy helps to stabilise
it behind the aircraft. But it’s still very difficult
for a pilot to manoeuvre the nozzle into the reception
coupling. Now it’d be easy enough to get
the probe into the basket, using just the robots alone. But what we’re trying to do is
to simulate the environment which the receiver aircraft
flies in. So we model the flight dynamics
of the two aircraft. We model the turbulence. We model the wake of the tanker.
and we model the bow wave on the receiver aircraft. And these simulations all
run in this box here. Those run in real time. And they pass the position
demands to the robot controller. So there are two computers
we need to look at in the controller. This is the main computer. And this is used to calculate
the motion trajectories for the robots. This computer here is the axis
computer, which is in charge of the motors. With the help of the Lund
University in Sweden, we’re intercepting the messages
between these two computers, which gives us direct access
to the motor controller. Effectively, we’ve hacked
the robot controller. The robots allow us
to manipulate the probe and the basket. So they follow the motions from
a real-time simulation, which runs in parallel. This simulation allows us to
develop the control algorithms for the real aircraft. And underpinning all of these
things is the idea of feedback control systems. So first, we’re going to take
a look at open-loop control systems, without any feedback. This is effectively a
remote-control helicopter. And we’re just going to look
at the elevation control. So what we’re controlling
is the fan speed. Now, it’s inherently
a stable systems. So I can let it go like this. And it’ll stay put. But if I give it a little
nudge, then it’ll start oscillating. And it’ll take awhile to settle
down to that stable position again. Now, we’re going to go ahead
and change that. We’re now going to close the
loop on that control system. And using the feedback, we get
a much more stable position. Now, if I nudge it, the fans
change speed to compensate. So that controller allows it to
stay in the same position. Now, what we’re going to try
doing is introducing a delay to this system. So we’re putting a delay
into that loop. And what that means is that by
the time the system is aware of where it is, it’s already
gone past the position it wants to be. And so it doesn’t have
time to compensate. So what you’ll see is that,
each time it goes past the stable position, the fans change
to speed too late. And it diverges from the
stable position. So that’s why delay is a bad
thing in a control system and why we’re trying to get
rid of it in robots. Let’s have a look at
what we just saw. We had an input going
into a controller. The controller determined the
fan speed for the helicopter. And then, the outputs from that
system were the position and velocity. So that’s a simple,
open-loop system. What we then did was we closed
the loop, by measuring the position of sensors and feeding
that back into the controller. What you saw was the effect. If we had a delay in
the sensors, the aircraft became unstable. Now, this is very similar
to the system we’re looking at behind us. We can split this
into two halves. The bottom half is
the real world. So we’ve got the physical
sensors that we’re hoping to evaluate. And the top half is completely
simulated. And on the boundary
between these two, you have the robots. Now, the robots translate the
positions from the simulation into real physical positions
that the sensors can then use to detect. Now, by bringing the robots
in, we now have two sources of delay. So there’s delays
in the sensors. And there’s delays
in the robots. We’re trying to evaluate the
delays in the sensors and how they affect our control
algorithms. So we need to eliminate the
delays from the robots, to make sure they don’t add
in to the effect. Now that we’ve ironed out the
creases in the test rig, the real work can begin.

Leave a Reply

Your email address will not be published. Required fields are marked *