Or just use wheels / a wheel. This whole humanoid thing strikes me as an addiction to old sci Fi stories.
Hydraulic systems have very little "give", unless you put a hydraulic accumulator (an air tank with a fluid/air barrier) in the system. Electric motors have plenty of "give". Forcing a motor to turn backwards won't hurt it. The gear train is usually the weak point. As motors and controllers have improved, robot gear reduction ratios have decreased, which reduces the load on the gear train and lets the motor absorb shock loads. Direct drive robots eliminate the gear train entirely. Here's a nice one.[1] "You cannot strip the teeth of a magnetic field" - General Electric electric locomotive rep, around 1900.
With modern motors, you can get huge torque with light weight, and cooling becomes the limitation. Schaft used water-cooled motors in their direct-drive robot. Google bought Schaft, ran them into the ground and killed them.
[1] https://shop.directdrive.com/products/diablo-world-s-first-d...
I'm still mostly convinced that harvesting the energy and re-using it ala elastic tendons is a decently good idea. But probably far too complex.
Variable compliance muscles are desirable, but hard to do. A pneumatic cylinder with adjustable pressure on both sides will do it, and Festo builds a lot of that for industrial automation. Two opposed springs pulled on by two positional actuators will do it, but that's kind of bulky. There's a hack called a "series elastic actuator", which is a rigid positional actuator with a stiff spring on the end. When it gets some pushback, the spring compresses, and the motor frantically tries to move the positional actuator before the spring bottoms out. This allows you to simulate a spring with off the shelf screw jacks.
Those new direct-drive motors are a good solution. Direct-drive pancake motors have been around for a while, but they used to be about a foot across. Now they're smaller. Probably a spinoff of drone motor technology.
One of the cheapest low-gear-ratios I found was the RMD series: (which is 1:6 or 1:10 gear ratio).
https://ozrobotics.com/shop/myactuator-rmd-x4-16-micro-motor...
This is one of those 80/20 things that is just glaringly obvious. Like lvl 5 autonomous cars vs lvl 3-4.
If the gear ratio on these motors is high, then there can only be faked compliance in the tuned force-torque controllers you mentioned. MIT's little cheetah robot, on the other hand, deliberately used low-gear ratios to keep things naturally squishy if needed. This is the way to go; putting elastic tendons or spring elements seems like a good idea but then you can't actually model the non-linearity well (the 1st order motor becomes a 2nd or higher order system).