A recent survey of "QDD" BLDC manufacturers and models

[Alan Timm]

I learned a new acronym today.  QDD == Quasi Direct Drive.

Its this new generation of low gear reduction integrated actuators.

Per the sheet I see 10 companies and 152 models.

I've bookmarked this for reference.  This is amazing!  I didn't konw there were so many.

I think the strongest that I found was 340nm.

https://docs.google.com/spreadsheets/d/1shXSUE_LMXkp5jD5uSPLa-V1LpkE5tsNzz1r5Yb0H7o/edit?gid=0#gid=0

Pulled from this youtube video:

https://www.youtube.com/watch?v=M78deObfLUE

[Chris Albertson]

Here is one more Open Source Humanoid robot.  Maybe even lower cost.   I think what it shows is that if you do use plastic for the structure, the size blows up. But they did that well and used the structure as “skin” to make it look more human-shaped.  And they reduced the size to what 3D plastic can handle.     I think this robot is maybe buildable by a normal person.      It has the lowest cost I’ve seen to date, but also the lowest performance.

But I think they have found the lowest cost humanoid.   But look, again it is remotely controlled. 

QDD, the idea behind the MIT Cheetah’s, involves low gear reduction (6:1 to 9:1) to enable motor back-driving. When commanded with limited torque, it creates a spring-like joint, enabling modern walking robots. Before this, they walked like robots and were fragile and couldn’t run or jump.

The above humanoid uses cyclical reductions, resulting in a “robot walk” with short, choppy steps. Cheetah runs like a cheetah, inspiring modern humanoid designs.  But these drives are printable, reducing cost.

You can DIY a QDD motor by buying a $20 Drove motor and 3D printing a double belt gear reduction. You’ll need two belts, a ball bearing, and metal shafts, but the rest can be printed. This works for a robo-dog side robot but fails for a humanoid due to its greater power requirements.    

The above Berkeley robot is a compromise for cost; they give away a little of the QDD effect and reduce the size to the point where you no longer need CNC metal parts, but you get what you pay for, and this little ‘bot would be a good platform for vision and planning research.

[Alan Timm]

Correction:  The strongest on the list is 450nm ?!?!?!

https://www.aliexpress.us/item/3256808430553578.html?gatewayAdapt=glo2usa4itemAdapt

Holy Moly!  Have you Visa or MC handy, it'll cost ya.

[Chris Albertson]

This chart explains humanoid cost tradeoffs. (Source https://arxiv.org/html/2504.17249v1 )

The black dots represent open-source humanoids. Compare the charge and budget to select one. Unitree appears to offer a balanced cost and performance.

[Alan Timm]

Yeah, and I think the berkeley bot shows the limitations of a 3d printed chassis.  I still think you can cheat a bit with a decent CF-PPA but your point still stands.  aluminum is stronger than 3d printed plastics for any given volume.

There's a really neat 3d printed BD-X project, and even at that smaller scale he had to move to a CNC aluminum frame.  I'll make a post about it later today.

On Tuesday, May 26, 2026 at 1:58:14 PM UTC-7 albertson.chris wrote:

[Chris Albertson]

The price is right, under $60 is cheap.  But 

(1) It is a 20:1 reduction.  Not at all backdrivable or “QDD” and

(2) it is dead slow.  In robot terms is it “fall you face” slow.

You might think it can move 90 degrees in 0.4 seconds, that is fast enough for a knee or hip  joint.  In terms of gross motion this is about right but what we care a lot about if we are balancing is “control bandwidth”.  How quickly can we make very tiny micro-motions to maintain balance.

The robot is falling, It needs to generate a force in the direction opposite.  We don’t need to move 180 degrees in .8 second, we need to move 1 degree in 2 milliseconds that is an 500 degrees per second. Yikes!    Numbers are made up but you can see balance drives the speed requirement

As fun as it is to just build something, Humanoids are like airplanes in that you like to calculate and simulate them to death before you start cutting metal.      Think for example if a robot walking outdoors steps on a 6mm tall rock and the contact point is on the outside of the left foot.   This is not even close to a worst case, I’d call it the normal case.   OK so the ground contact is by surprise now 25mm to the left of expected.  The robot is 60Kg and now you have a 25mm x 60Kg sideways torque from gravity.   The IMU senses this and the PID balance loop commands a restoring torque from the leg motors.      

With my Quadruped, the above was solved by using balls for feet.  With spherical feet, the contact point is never a surprise.  The contact is ALWAYS a line between the two feet on the ground.  Then there is one 1-DOF balance problem.   Quads are much easier. 

Finally, speed of gross motion DOES matter.   A robot, if it is to run, needs to have enough speed and power that it can leave the ground by pushing with one foot.  The definition of “run” is that there is a time when all feet are in the air.     This little robot seems to be able to run.  How much vertical velocity one needs to leave the ground for half the stride period?

I don’t mean to be pessimistic.  All this is possible, but unlucky with $60 motors on a full size humanoid. 

That said, for a quadruped, here are my favorite motors, although you need to add belt reduction and a controller.  These have enormous control bandwidth because you can put up to 50 amps through them for a millisecond or so as long as the average temperature is OK; they work.  You can move that 1 degree in a millisecond.  I have one on a test stand, not even close to in a real robot yet.