This is what I sent to the Popular Science reporter (I also talked to
him, which is where some of the quotes came from).

I have 3 types of comments.

1) High level:
This is a big step forward for reinforcement learning, bridging the
gap between machine learning and state-of-the-art robot control. The
work shows how to use what we already know, and learn what needs to be
learned. This makes learning faster and more reliable.

2) Nit-pickingly technical:
One reason the proposed approach works well is that actuator dynamics
are decoupled (the left elbow doesn't affect how the right shoulder
works).  Rigid body dyanamics are coupled, so one has to take into
account everything to predict anything. It is easy to learn things
that are decoupled or isolated phenomena. It is very hard
(exponentially more so) to learn about coupled phenomena, like the
weather. So what happens if coupled phenomena have to be learned. For
example, I would love to see this applied to more deformable (and thus
less expensive but more complex to model) robots. Anymal looks like a
very stiff box with relatively lightweight legs (that can be viewed as
force sources due to the difference in time constants). This means the
significant dynamics are those of a box, which are pretty simple. What
happens when this approach is applied to state of the art humanoid
robots, that tend to have lightweight frames that flex during behavior
(as human bones, mass distributions, and joint structures do in fast
movement). Rigid body dynamics are now violated, since real
(lightweight) robots aren't rigid, might have varying mass
distributions as liquids slosh and wires flex, and don't necessarily
have well defined joints. Lots of folks have built robots that match
rigid body assumptions, but they typically spend a lot of money doing
it (multiple millions for Boston Dynamics humanoids). This is due to
high quality bearings, exotic materials, expensive valves or motors,
etc. Widely used robots in the future might be cheap, and probably not
very rigid. I am trying to build balloon robots, and also think
Nerf(TM) robots would be a great idea.

There is great video out there about how human tissue jiggles, bones
bend, and knees move (out of plane) when you run. I am trying to find
Youtube versions.  Your head isn't rigid (you don't get to see your
brain compressing and sloshing around).
https://www.youtube.com/watch?v=On1CsbTwlDs 2:35 for example
https://www.youtube.com/watch?v=hORwtZ2vLuw
Here are some quadrupeds, to compare to ANYmal.
There is nothing rigid about this motion (except each individual bone)
https://www.youtube.com/watch?v=v7p6VZiRInQ start at 1:00
https://www.youtube.com/watch?v=V8vejjVgIHg look for high speed video
https://www.youtube.com/watch?v=Ldz2eB2zSvY

3) Joke:
Now that there is video evidence that these researchers abuse their robot,
AND they programmed their robot to learn, these people will be the first
to go in the ANYmal-led robot revolution.