Roboy used dyneema tendons if I recall correctly. Fluidic actuators is another option. IMO, additional sensors and sensor fusion are necessary but this will raise the costs and demand to control software significantly. We are researching humanoid robots for quite some decades now and these problems are easily underestimated (similarly to autonomous driving). I doubt we'll see them in our houses very soon.

The issue is not that these problems are easily underestimated, but that the researchers are very proficient at repeating the same mistakes over and over again.

Or both …

I think you're misunderstanding the essence of the problem. If you use tendons, you'll need a neural network in your control loop that can learn continuously so that it compensates rope stretch and changes in friction through wear and tear.

A lot of problems in robotics reduce down to continual learning. Essentially all system identification tasks become obsolete the moment you have a self learning system and yet we have an AI industry preaching that AGI is around the corner without this "crutch".

Not to mention that they easily break and are annoyingly tedious to be replaced (I had to assist often enough on such "surgeries" in our lab)

You don't necessarily need neural networks for that, there are more specialized function approximations that learn faster and enable life-long learning for kinematics/dynamics and extrapolate better than NN which are more general purpose.

You could add encoder patterns to the tendons, like stripes. And then use something like a mouse-sensor to track them. This is still much lighter than servos.

Another idea is to use an external camera or two and to track the fingers with a deep learning model. But this can become messy if other objects are in view. And it might also introduce more control delay in the feedback loop than a simple sensor.

If space permits you should add absolute encoders in the joints. In a finger that’s likely a challenge. In an arm, that’s easy

I would think pressure and angle sensors tied to a PID loop would suffice, but I'm not a roboticist.

If you can get a feedback loop at the respective joint, then yes, this might be enough. In case of an arm (not this hand), you can often only observe the end of a chain, and then, that’s more complex