There's a basic conflict. Small electric motors want to turn fast, so they're usually followed by gear reduction. But that loses feedback precision and back-driveability. A pure direct drive motor works great, but they're large diameter devices. Some SCARA robots use them, but the motor is a foot across. Washing machines have gone direct drive, since there's enough space for a large diameter motor. There's a direct drive electric motorcycle with a hollow rear wheel. There are "pancake" motors, with large diameter but little thickness. None of those devices have a good form factor for humanoid robots.
That leads to tradeoffs such as quasi-direct drive, where there's some gear reduction, but not too much. The article suggests that 20:1 is an upper limit for back driveability. That's pushing it for a leadscrew-type device, but maybe it's possible now.
It's neat seeing all this progress in robotic components. Historically, robotics has been a small niche, and had to use components developed for other purposes. This made for clunky robots. Now we're seeing more purpose-designed components made in volume. Drones made 3-phase synchronous motors and their controllers small, light, and cheap. Now the same thing is happening for other needed components.
Looks like, when the AI guys get their act together on manipulation, the machinery will be ready.