I've been investigating the possibility of using a worm gearbox with encoder as a more economical substitute for servos, even for things like Robo-One humanoids. It sounds nutty, but it just might work!
Let's start with the Tamiya Worm Gear Box HE  (about $10). This can be assembled with two gear ratios, 216:1 and 336:1. It comes with a Mabuchi RE-260 motor, which gives us roughly:
No gearing: 15 g cm of torque at 8300 RPM 216:1 gear: 3240 g cm of torque at 38.4 RPM 336:1 gear: 5040 g cm of torque at 24.7 RPM
So we're over 5 kg cm of torque right out of the box. But Mabuchi makes a variety of stronger motors  which are only slightly larger; I'm sure we could use them anyway since the motor doesn't really sit inside this gearbox. There's the RE-280, which is a bit faster and stronger, and then there's the RE-360, which is over twice as strong:
No gearing: 32 g cm of torque at 8321 RPM 216:1 gear: 6912 g cm of torque at 38.5 RPM 336:1 gear: 10752 g cm of torque at 24.8 RPM
Now we've got over 10 kg cm of torque, which is what I've been told is necessary for at least some of the joints in a humanoid robot (probably the legs and hips in particular). Compare this to the Hitec HS-5945MG digital servo, which generates 9200 g cm of torque, and has a speed equivalent to 62.5 RPM. Yes, that's over twice as fast -- but it also costs $90.
I haven't found a source yet for the RE-360, but the RE-260 costs about $2, and the RE-280 (the next step down from the RE-360) is $2.25, so I doubt the RE-360 costs more than $5 or so. We're at about $15 for the gearbox and better motor. (Note that this motor was the result of a pretty quick search; more searching might turn up a faster one, to make up that lost speed without losing torque.)
Of course, we still need an encoder scheme; this can be done with a quadrature encoder and an additional optical interrupt switch , which add a couple more bucks to the total.
It's also possible that you could ditch the quadrature encoder and use another simple opto switch, under the assumption that you know which way you're driving the motor, as at least one guy has done . Since this uses a worm gear, it can't be driven from the outside, and if any errors do creep in, you can clear them every time you cross your zero point (which you detect via the other opto switch).
You would of course still need a motor controller. Something custom to this application might be appropriate, especially if you're planning a bot with 20 of these things.
So to me, it looks like for under $25 or so, you could get performance comparable to a $90 servo, and have the added benefit of a very convenient form factor and a complete through-shaft for making nice balanced joints. (In fact, I suspect that you could get by with less torque because you're driving a joint axle, instead of only driving one side of a joint as most servos do.)
What do you think? Am I off my rocker?
- Joe