Microduck

[Stephen Williams]

Looks very fun, while being minimal, clever, and with personality.  In addition to a Humanoid Scale, we could use an Animoid Scale to rate & rank robots.

Animoid Scale attributes that Microduck has:

• Head & neck
• Large range of motion up, down, turning.
• Has a functional mouth: dog / duck
• Has expressive sounds
• Has camera eyes
• Hears
• Legs & locomotion
• Can walk or waddle
• Resilient control capable
• Can recover from falls, out of balance

https://makethings.make.co/p/walks-like-a-microduck

https://www.youtube.com/watch?v=aG5BV0e-GrQ

Said to be open source, reinforcement learning capable.  Seems like a nice hobbyist platform.

Stephen

Previously, I wrote about Reachy, a desktop robot from Pollen Robotics in association with HuggingFace. That partnership has come up with a new kind of open-source robot — the Microduck. “At 25 cm tall, 14 cm wide and under 800 g, Microduck is built for the messy reality of reinforcement learning,” says the press release.

Reinforcement learning is a way that robots can learn to do something instead of being taught or programmed to do it. The key is that they can try a million ways until finding something that works. It’s like a child learning to walk and falling all the time until they figure it out. They call this kind of learned behavior a policy. They explain that doing reinforcement learning on the robot itself is slow. Their solution is to “train the robot in simulation and transfer the learned policy to the real world, known as sim-to-real transfer.”

A typical development loop looks like this: train in simulation (on your computer), deploy on the real-world robot, observe the behavior, refine the simulation parameters, repeat the process, share the policy with the community.

The results looks like a lot of fun but it should be fun for developers to train. There’s something a lot more playful with a robot that looks like an animal rather than a human. Maybe there are lessons to be learned from cartoon characters, the world of animation is a lot more fun and playful rather than “realistic” characters.

Stephen

--

Stephen D. Williams Founder: VolksDroid, Blue Scholar Foundation 650-450-8649 | fax:703-995-0407 | s...@lg.net | https://VolksDroid.org | https://BlueScholar.org | https://sdw.st/in

[Stephen Williams]

$399 preorders: https://pollen-robotics.com/microduck/

https://github.com/pollen-robotics/microduck

The hardware does not seem to be open source.  Maybe that doesn't matter as much or all of it may be obvious, easy enough to reproduce & hack.

This much is open source:

The SDK, the simulation and the full RL training stack are on GitHub. What the robot runs is what you can read, fork and retrain.

Apache-2.0

The whole software stack, permissively licensed

Stephen

[Michael Wimble]

Steve, I think you should first make one of the MicroDucks in a humanoid size. This will be particularly useful for:

* Halloween

* Ridding your lawn of unwanted kittens

* Flattening boxes for the trash

and a few things I’m forgetting

--
You received this message because you are subscribed to the Google Groups "HomeBrew Robotics Club" group.
To unsubscribe from this group and stop receiving emails from it, send an email to hbrobotics+...@googlegroups.com.
To view this discussion visit https://groups.google.com/d/msgid/hbrobotics/09dffdd4-53ee-4bb9-ad89-0c23befdc9ad%40lig.net.

[Chris Albertson]

Roller skating?   That is a truly original idea, and look how well it works.  So good to see something new like this.    

I think the days of writing motion generation by hand are over.  Today we write “loss functions” and design network topologies.  Got to learn this, or risk being a dinosaur.

You know what I don’t see here?   Why aren’t people training ROS-based "smart carts”?    Lots of people have built a differential-drive wheeled platform with sensors and then connected that to ROS-based SLAM.     The EXACT same hardware could be run by an RL-trained network.   Nothing new to build.  Repurpose the existing hardware. You would train a pixels-to-motors network to go to various locations and not run into stuff along the way.    I think better to do that than try to build and train a housekeeper to clean a kitchen using arms and hands.   This level of AI can run even on low-cost hardware.    

.

--
You received this message because you are subscribed to the Google Groups "HomeBrew Robotics Club" group.
To unsubscribe from this group and stop receiving emails from it, send an email to hbrobotics+...@googlegroups.com.

To view this discussion visit https://groups.google.com/d/msgid/hbrobotics/21aa8684-a155-43dd-b979-aab46c14cff5%40lig.net.

[Chris Albertson]

Hint: Read the GitHub.  The software design is good enough to warrant some study.   The robot has a fixed set of about a half dozen actions it can do like “kick with left foot” or “stand” and some others.  You could in theory teach it more.  You would start by creating a training environment, training the model, then modifying the real-time source code (written in Rust) to handle the new skill.    For safety, the robot CANNOT be given raw motor commands or joint angles.  You can only tell it to do a trained skill.   The robot software is in Rust and is easy to read with many comments.   The training environment is all Python and Mucojo.   If adding a new skill 90% of the work is in the training environment.

I think this robot could be trained to walk up very modest stairs or do a spin in place on skates.  To do that, you have to think of some kind of training method and write the software to compute a reward function and then integrate to new model with the real-time software.   You would learn quite a lot that could be applied to other robots.   You need a good Nvidia-based training computer of rent time on a Hugging Face server (they tell you how to do each of these)  

Reading all this I can see the author dealt with the usual problems with sim-2-real.   It seems that to make sim-2-real work you need a very accurate simulation.  Even things like s"lop” in the joints the gear backlash are modeled in the simulation

I guess what I am getting at is that the designer clearly intended to have people peeking under the hood

I don’t want to describe it all here, just want to motivate someone to read the code.    This is how rotobots will work going forward.  I think we have turned the corner away from hand-coded motion control.

On Aug 29, 2026, at 1:00 PM, 'Stephen Williams' via HomeBrew Robotics Club <hbrob...@googlegroups.com> wrote:

Looks very fun, while being minimal, clever, and with personality.  In addition to a Humanoid Scale, we could use an Animoid Scale to rate & rank robots.

Yes you could, but changing even the slightest measurement of the robot would require a compete retraining of the model using your modified robot.

[Stephen Williams]

Another cute little robot, no price I could find yet.

https://www.hightorquerobotics.com/

Stephen

[Kai Mai]

Hi everyone,

This thread's already touching on a lot of what's relevant to Quaddle, my company's newest open-source robot — cute little robots with personality, sim2real, movement training. I just posted about it on a new thread, so I won't repeat the specs here.

On the movement-programming discussion — Chris, your point about writing loss functions and designing new topologies got me thinking about the other end of that spectrum. On Quaddle Buddy and Scout, every joint is a position-feedback servo, so instead of training anything, you can just physically grab a leg and walk it through a motion, and it's captured directly as a replayable behavior. No reward shaping, no sim — just teach-and-replay. 

I'd love to hear everyone's feedback.

Best,
Kai Mai