Robot arms should use harmonic drive gearboxes and servos, IMO. In practice, if the loads are not massively high, the important thing is closed loop steppers or servos, with accurate homing routines and good error handling.
It depends on the kind of load and the job. Harmonic drives tend to shear teeth if they get a shock load. It's good to have a back-driveable gear train, so that if the arm hits something or is overloaded, the forces push back to the motor, and the controller can see them and stop or deal with the problem. For something like a pick and place machine, you don't need that, but if you're trying to put screws into holes and might miss, you do.
3-phase servomotors (which used to be called "brushless DC servomotors") are much more available than they used to be. Drone motors are tiny 3-phase motors, and can be controlled as servomotors. The drone industry has done a lot to make such motors cheaper.
The controllers are much smaller and cheaper now, too. They used to be the size of a book or worse. But they're still more expensive than they should be. I was once at a trade show talking to a rep from a motor company I'd used, and noticed they now sold controllers, too. He said, yes, they had to get into the business because others were selling controllers for 10x the cost of the motor, and they cost about the same to make.
For a long time, you had to bolt the encoder on the back of the motor, and encoders cost way too much. Motors with built-in encoders have become more common.
"RC servos" from model aircraft are cheap, but crappy actuators. They're advancing from 1970s pulse width modulation control to 1980s serial, and for under $20 you can now get some force feedback.
When we build actuators in the Leg Lab, every actuator had some electrical feedback element. For example, leg length actuator had a resistive strip that was as long as the leg's travel. A fixed 'wiper' allowed us to therefore have a potentiometer, giving a leg-length-proportional voltage on the wiper when top and bottom were driven by a constant current source. This was used in a 500Hz (sometimes 1 kHz) servo loop for position control.
(somewhat related): We also had a velocity sensor on the legs -- a simple coil of wire and a magnet (but COTS). This gave us a Velocity signal directly -- rather than differentiating the position measurements. Velocity estimates made by differencing position measurements tend to be noisy, and filtering esp FIR filters introduces phase delay -- things you do not want. Lesson was: If at all possible, get a sensor that senses the quantity you want to sense, so you do not have to integrate or differentiate to get your desired quantity. It's the same reason we had a REAL gyroscope, and not a rate gyro (as in most MEMs IMUs -- in IMUs they call them 'gyroscopes' but they are actually 'rate gyroscopes' -- they only directly measure "theta-dot").
> Velocity estimates made by differencing position measurements tend to be noisy, and filtering esp FIR filters introduces phase delay
Yes. Which is why you see encoders for sale with a zillion points per rev. Not because the position is that precise, but so that you get enough events per second to calculate velocity. Actual tachometers, which are little DC generators, are sometimes used in servo systems, but not often. The last time I saw one, it was on a mainframe tape drive.
You do want velocity and force feedback if you're doing anything beyond pick and place in a totally controlled environment. Simple preprogrammed blind movement is most of industrial robotics, though. There are fancier systems, but most of them don't sell, because the engineering cost exceeds the value add. Sometimes you see a little force feedback to get something inserted properly, but more often you see some mechanical spring setup to get the same result.
I used to be into legged locomotion, but there's no profitable market. Many of the technical problems have been solved, though. The sensors and actuators are good enough. Electric motors finally are strong enough. (It's sad that Schaft went under. Google bought them and dumped them.)
Robot arms should use harmonic drive gearboxes and servos, IMO. In practice, if the loads are not massively high, the important thing is closed loop steppers or servos, with accurate homing routines and good error handling.