What are the tradeoffs in building dexterous robotic hands? To answer this question, let us examine the two main approaches to building them:
Tendon-driven hands, with remote motors driving finger joints using tendons. Some notable examples of this approach are Tesla’s Optima, Shadow Robot, 1x NEO, and Kyber Labs.
Micromotor-driven hands, with small motors and high gear ratios on every joint. These have been used by Covvi, Quji, Figure AI and Psyonic.

These two approaches lead to different engineering tradeoffs, along the following dimensions:
Payload: How much weight can the robot carry?
Heating: How much does the hand heat up, when stationary and when in motion?
Sim-to-real gap: Does robot training during simulations translate to the real world?
Degrees of freedom: Does the hand have sufficient DoF to perform “routine” tasks?
Speed of movement: How quickly can the robotic hand perform tasks compared to its human counterpart?

Here is how the two mainstream approaches address these tradeoffs:

Vision-free MIT Cheetah | YouTube

Proprioceptive actuation greatly reduced the reliance on complex sensors, and allowed the MIT Cheetah to perform fast, dynamic movements, such as running at up to 13 mph, jumping obstacles and walking on gravel. Today, this technology shows up in the form of dancing robots.

The use of high-torque motors with low gear ratios (the 2014 paper used 1:5.8) gave this approach its name — quasi-direct drives.

This paradigm is well-suited to legs, because the lower body can accommodate a large and heavy motor. However, because electromagnetic motors dramatically lose torque with size, it cannot be applied to robotic hands.

If only there was an actuator that had 100x the torque density of electromagnetic motors.

[1]  The MIT Cheetah, an Electrically-Powered Quadrupedal Robot for High-speed Running, 2014: one of the early publications from Dr. Kim’s group.