I don't see why electric motors are not backdriveable enough. Boston Dynamics' Spot uses electric motors with harmonic drive gearboxes (or some other backdriveable transmission) as an example of this. Similarly, series elastic actuators are very capable of giving the characteristics you deem necessary to work alongside humans albeit with lower bandwidth than the harmonic drive gearboxes.
You’re spot on with the harmonic drive motors. They do have good backdriving and low backlash, which is great. I certainly don’t mean to say that there is no place for electric servo motors in robots, they are fantastic for many applications, especially if you need precision and accuracy in movement. And the Spot (along with many copycat robots) make great use of them.
But the problem remains that there is still a severe trade off between speed, force and backdriving. The Spot really doesn’t have that high of a max force it can generate, and no where near that of a similarly sized dog. Additionally, the motion produced by the kinematics system it uses it relatively simple with far fewer degrees of freedom than a real dog.
The general consensus in the field is that if you want to have a robot which can do everything a human can do, you need to at least match the degrees of freedom a human has in motion. It would be very difficult to fit 100+ harmonic drive servos into a robot. Of course controlling all of this effectively is extremely difficult and that’s often why robots like spot use vastly simplified kinematic models, because it makes the equations go from near impossible to solve to just hard.
series elastic actuators are promising and I have high hopes for their use in the future. I have some colleagues who work with them. The limitations there are more to do with complexity and finding ways to dynamically tune the force impedance of the springs.
As someone with a similar background, I believe some of the confusion is because there is a lot of overlap. System identification is very similar to supervised learning, however there are other learning "methods" that still fall under the umbrella of ML/AI. For example, unsupervised learning doesn't really have a good controls analog (as far as I know). Reinforcement learning on the other hand is somewhat analogous to model predictive control.
A better way of phrasing your point is that ML/AI is "just" optimization.
I don't know about other languages, but in Julia I've heard people often say that loops end up faster than the equivalent vectorized code. So while this is true for Python/Matlab I don't think it is good universal advice.
That said, matrix notation can sometimes be the more readable way of expressing a calculation.
I might have used an early beta of Julia circa 2018 or something, but the chorus that it performs like $static_fastlang doesn't match the experience I had.
I'd don't know what you mean by "performs like" but it's definitely significantly closer in runtime speed after the first run to compiled languages than interpreted (in particular Python/Matlab).
Anyways, my point was more in response to this from the person I responded to: "my instinct is to look for opportunities to vectorize by rewriting loops into matrix notation on paper and then expressing them as array calculations". In some languages (Julia in particular) that is slower than the equivalent loop based code.
