Re: These ARE The Droids ...

[Alan Timm]

Sounds like fun!  I've personally been wondering if all of this RL training Sim 2 real stuff works equally well on servo based robots, especially since these serial bus servos are readily available.

Are you joining tomorrow?  I'd love to hear more about it.

Alan

On Wednesday, October 8, 2025 at 11:30:57 PM UTC-7 Chris Mayer wrote:

Dan Albert and Chris Mayer are starting up the second wave of building a full sized humanoid android.

We've done it before, and want to take what we've learned from our first sets of walking robots to the next level.

(Our temporary code name is "These ARE the droids you're looking for.")

(WATSON on the table, May 2016)  https://youtu.be/ChYZN1z0_mc?t=821

(Walking demo, July 2017) https://www.youtube.com/watch?v=BLBBoPJLEh0

Join in on the fun, especially if you could help with the following.

    Mechanical Engineering

    Robot Simulation

    Software Engineering

    Electrical Engineering

    Power Distribution

    3 Phase DC motor control

    3D printing

    CNC milling

(We're looking for active members, not arm-chair quarterbacks.)

We plan on crowdfunding once we get close to a working prototype.

If you might be interested in joining this project, please reply privately.

We're planning a kick off Zoom this Monday evening.

[Chris Albertson]

On Oct 10, 2025, at 12:31 PM, Alan Timm <gest...@gmail.com> wrote:

Sounds like fun!  I've personally been wondering if all of this RL training Sim 2 real stuff works equally well on servo based robots, especially since these serial bus servos are readily available.

Yes and no.   If you let the simulation play, then in theory, it could be moved to a servo-based system. But you quickly find that the servos lack the performance to actually execute the data that is sent.   It is not simply a lack of torque but a lack of bandwidth, acceleration, and control model.

The problem is fundamental. Servos are basically gear motors where the motor runs much faster than the output shaft, and there is a rotational encoder on the shaft.   The servo can be commanded to a location but CANNOT be commanded to a given torque and is not back-drivable.    This pretty much means that servos just cannot work for the kinds of robots we see on the Unitree r1 video or Boston Dynamics.      Servos can do slow, scripted motions but not cartwheels on a natural surface.

There is a reason ‘everyone’ is using BLDC motors with about a 6:1 or 9:1 gear reduction.   The discovery that low ratio reduction “works” was what changed walking robots.  It was a kind of revolution.

[Alan Timm]

Hey Chris,

I hear ya, but...

The Aloha system really kicked off this RL Training regime and showed us what's possible.  And those are all using bog-standard serial bus servos.

And even though all the attention is on expensive BLDC humanoids, the aloha folks are still hard at work, and we're even seeing some interesting developments with lerobot which is also serial bus servo based.

What I don't know is if anyone's tried using these tools for walking with serial bus servos whether quadrupeds or bipeds.

I'd say it's definitely something worth exploring.  And if it works it opens up advanced robotics without the advanced price tags.  Could be fun.  :-)

[Chris Albertson]

Yes, 100% true.  But Aloha is not walking on two legs.

My comments about servos were only that you cannot do dynamically balance walking with servos.   Of course, you can train a system with servos.

If you are building arms and hands, servos work very well.    But walking is different.  As you (or a humanoid robot) walk, we translate forward while the rear foot is fixed on the ground.   Eventually, our weight goes over the rear foot, and we fall.  But the other leg swings forward and breaks the fall.       The problem with servos is that we need to make very fast corrections while this happens, and you simply cannot push commands down to the servo fast enough to keep balance and (2) It is unlikely that you will find servos with a good enough power-to-weight ratio to support not just the weight of the rest of the robot but to propel it off the ground in a jump.   Then there is the problem of servos not being back drivable.   A robot needs “springs” in the legs for compliance with not-flat ground, and back drivable motors with constant torque make good springs.

The thing about humanoid walking is that much of the time we have only one foot on the ground and worse, our weight is not even over that foot.   

The Aloha system is trainable and is a MUCH better kind of project for a hobbyist to try because the cost can be very low.  You can build a servo arm for a little more than the cost of the servos

Sorry if I was not clear, servos are compatible with RL training, but servos are not compatible with dynamically balanced walking.

I even have a robot that uses servos and walks but lacks enough control bandwidth to balance itself.  You can see in the slo-mo video that I had to take baby steps so the robot would not have “time to fall” and this would never work on a uneven floor.  WhatI would like is for the 'botto be able to balance on two feet while the other two move much farther forward.  or for an actualy running gait where all feet leave the ground.   I need more powerfull motors

On Oct 10, 2025, at 9:17 PM, Alan Timm <gest...@gmail.com> wrote:

AlohaYes but Aloha is nt walking