Why almost no robot motor drives its joint directly
Small electric motors are naturally built to spin fast and produce relatively little torque, a typical DC motor might run at several thousand RPM while only outputting a fraction of a newton-meter of torque at that speed. Almost nothing a robot needs to do wants that combination: robot joints need to move at a few tens of RPM at most while producing torques that can be tens or hundreds of times higher. A gearbox closes that gap by trading speed for torque: a reduction ratio of N:1 divides output speed by N and, in an ideal lossless gearbox, multiplies output torque by that same factor N. This is why gear reduction, not motor selection alone, is often the deciding factor in whether an actuator can actually do its job. A small, cheap, fast-spinning motor paired with a well-chosen gearbox can outperform a much larger direct-drive motor for a fraction of the mass and cost.
Real gearboxes aren't perfectly lossless, though. Every meshing gear stage introduces friction, and that friction eats into the ideal torque multiplication, so actual torque gain is somewhat less than the reduction ratio suggests, and efficiency drops further as ratio and stage count increase. Different gear architectures make different tradeoffs here: simple spur gear trains are cheap and easy to inspect but bulky at high ratios, planetary gearboxes pack a high ratio into a compact, inline package with reasonably good efficiency, and worm gears achieve very high single-stage ratios but at ratios of 20:1 or more, are typically not backdrivable at all.