Also the sets made by Harmonic Drive (the company) are definitely designed to be back-drivable. Your conment made me smile, because I remember that I heard the same thing about these gears. My first week at Rethink I asked my boss about it while looking on in slight horror at him moving around an unpowered robot while he was explaining the system to me -- I was very concerned we were about to break something!
This is all much easier with three-phase brushless motors. You can hold the motor rotor at a specific angle under computer control.[1] With DC motors, you could just exert a torque, and it was hard to maintain tight position control, especially if the load changed. So with DC motors, robots needed more gearing down. The usual arrangement was a small, fast motor and a high-reduction geartrain. Adept Robotics was the first to go in the opposite direction with a direct-drive motor, a pancake-like thing almost a foot across.
(Robotics is full of routine engineering headaches like that. Many have been solved in recent years. Motor control used to be a much larger headache, and good motor controllers used to be a lot more expensive.)
[1] http://hackaday.com/2015/04/20/driving-a-brushless-dc-motor-...
As for gear ratio -- depends on which joint, but Harmonic Drive's standard gearboxes come in ranges of 50:1 to 320:1, and I'd say they were all within that range.
As far as direct drive motors go, all I can say is that I was very aware of them, and very interested in looking at it. (I was on the embedded systems team doing motor control), but by the time I joined, the architecture had already been decided and robot's were being built. That said, there are certain other difficulties there, both with the control electronics for larger motors, and with the relative cost of the motors, especially since Rethink is going for the goal of reducing robot costs. I can't say they won't go that way some day. I think at the very least, it's worth exploring, but I'm not longer there so I don't have much say in that!
Other fun stuff -- Even though we used a harmonic drive, it was coupled to what is basically a torsion spring between it and the output. This arrangement is called a series-elastic actuator, and it allows you to do some neat stuff with rapid torque sensing, both from the control and safety side of things. It's also a lot of fun for the control engineers trying to control something sitting at the end of a spring :-P
I'm not trying to say it's impossible, or to knock anyones project or sounds argumentative. It's been done, as you pointed out! Just wanted to point out that while in theory it's much nicer to be able to do direct torque control on the motors, the headaches to become apparent in practice. (I still love the idea of direct drive :-D )
A double-ended pneumatic cylinder will do this. Two motors pulling on opposed springs will do this. A series elastic actuator can fake it by active control. You don't get the energy recovery that way, though. There have been systems with wires and pulleys to get muscle-like actuation, but they're bulky.[1]
The series elastic actuator is a position actuator driving a stiff spring, with positional and force sensing at both ends of the spring. If a load starts to compress the spring, the control system frantically runs the actuator to unload the spring before it compresses much. This gives the illusion of a less stiff spring. The spring stiffness simulated can be adjusted in the control system.