Unitree G1 Humanoid for $16k?!?!?! Yes Please!

198 views
Skip to first unread message

Alan Timm

unread,
May 13, 2024, 3:02:59 PM5/13/24
to RSSC-List
https://www.unitree.com/g1/

Very exciting!  There's one small disclaimer at the very end of the video:

"* Different models have very different configurations, functions, and prices of robots.  Please pay attention to the distinction."

unitree g1.jpg

Chris Albertson

unread,
May 13, 2024, 4:07:55 PM5/13/24
to gestalt73, RSSC-list
$16K is for the base model, The video is of the more expensive version.  But still, $16K is a fantastic price.   I have said in the past that when these robots will sell for the price of a car, they will be popular.  It looks like that has happened.   (Why a car price?  Because cars are personally affordable to many a billion or more people)

I just wrote in another email that what’s next is to learn how to connect GPT-4 with a physical robot.  Not so the robot can say “I like kittens” but so that GPT-4 can control the motors and sensors.  




<unitree g1.jpg>


--
You received this message because you are subscribed to the Google Groups "RSSC-List" group.
To unsubscribe from this group and stop receiving emails from it, send an email to rssc-list+...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/rssc-list/5a8d2953-0950-448d-9488-d2481857b8fdn%40googlegroups.com.
<unitree g1.jpg>

Gmail

unread,
May 13, 2024, 5:04:58 PM5/13/24
to Chris Albertson, gestalt73, RSSC-list
“ learn how to connect GPT-4 with a physical robot.  Not so the robot can say “I like kittens” but so that GPT-4 can control the motors and sensors.  “

Yes, I am now working on that !



Thomas

-
Want to learn more about ROBOTS?








Alan Timm

unread,
May 13, 2024, 5:58:39 PM5/13/24
to RSSC-List
Yeah,

Looking at the limited initial info available, they're attempting to segment their product line just like they did with the Go1 and the Go2.

Aside from the usual silliness locking advanced software features to the higher cost model, the base model does not come with hands?  That seems odd.

Also there's no mention of GPU acceleration which is weird because everything including this humanoid is running deep learning.

I can't wait to see more details about this, from everything I know about the state of hardware this should be at least 10-20x the cost they're advertising.

My guess?  They're hiding a significant portion of the cost behind some type of required cloud subscription service.

The Aldebaran guys did this with the Pepper.  Only $2400 for the robot, but requires a $600/m subscription, minimum 3 years.


g1 differences.jpg

Chris Albertson

unread,
May 15, 2024, 1:11:01 PM5/15/24
to gestalt73, RSSC-list

On May 13, 2024, at 2:58 PM, Alan Timm <gest...@gmail.com> wrote:

Yeah,

Looking at the limited initial info available, they're attempting to segment their product line just like they did with the Go1 and the Go2.

Aside from the usual silliness locking advanced software features to the higher cost model,

The reason is that the low-priced version comes with a cheaper computer, When you pay for the programmable version you get a much more powerful computer.  An Nvidia Jetson “Orin” can cost $2,500


the base model does not come with hands?  That seems odd.

It is common to sell robots with no hands (or “end effectors”) because they don’t know what task the robot will do and usually you need to select hands that are suited for the specific task.    KIt is kind of like why bicycles are sold with no pedals, everyone has their own idea of which is best


Also there's no mention of GPU acceleration which is weird because everything including this humanoid is running deep learning.


Are you sure any of the software uses deep learning?  I did not see anything about deep learning.   It is rare to control robots with any kind of “AI”.   Mostly they use very sophisticated classic control theory.   BD’s “Atlas is “whole body MPC”


I can't wait to see more details about this, from everything I know about the state of hardware this should be at least 10-20x the cost they're advertising.

Decent actuators cost maybe $250 each.  A humanoid is about 28 degrees of freedom.  Then add structure and batteries.   If the robot is mass produced it should cost maybe 20K.    But if you only build one robot, then millions.


I know someone who’s company where he works bought a Boston Dynamics “Spot”.   BD sells them for $70k.  But they paid $200k.  Why?   Basically for software and custom programming.  The basic Spot is very dumb.   You drive it around with a remote controller it stays upright and avoids obstacles, that is it.  It is a $70,0000 toy.    But they wanted a robot that could do plant inspections.   They had to buy some sensors and some software from a third party consulting company.

Humanoids will be the same way, usless without added software.  The software will need to be custom to your application and therfor expensive.   Software is at least $100 per hour and you can expect large projects to be at least in the tens of man-years of labor.   I did a few cost estimates where I used to work and at frst I was shocked to see how little $1 million buys you.




My guess?  They're hiding a significant portion of the cost behind some type of required cloud subscription service.

The Aldebaran guys did this with the Pepper.  Only $2400 for the robot, but requires a $600/m subscription, minimum 3 years.




<g1 differences.jpg>

On Monday, May 13, 2024 at 1:07:55 PM UTC-7 albertson.chris wrote:
$16K is for the base model, The video is of the more expensive version.  But still, $16K is a fantastic price.   I have said in the past that when these robots will sell for the price of a car, they will be popular.  It looks like that has happened.   (Why a car price?  Because cars are personally affordable to many a billion or more people)

I just wrote in another email that what’s next is to learn how to connect GPT-4 with a physical robot.  Not so the robot can say “I like kittens” but so that GPT-4 can control the motors and sensors.  




On May 13, 2024, at 12:02 PM, Alan Timm <gest...@gmail.com> wrote:

https://www.unitree.com/g1/

Very exciting!  There's one small disclaimer at the very end of the video:

"* Different models have very different configurations, functions, and prices of robots.  Please pay attention to the distinction."

<unitree g1.jpg>


--
You received this message because you are subscribed to the Google Groups "RSSC-List" group.
To unsubscribe from this group and stop receiving emails from it, send an email to rssc-list+...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/rssc-list/5a8d2953-0950-448d-9488-d2481857b8fdn%40googlegroups.com.
<unitree g1.jpg>


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

Alan Timm

unread,
May 15, 2024, 3:43:09 PM5/15/24
to RSSC-List
Unitree G1 Video from ICRA2024

This appears to be the $16k base model with fixed position hands and 1dof waist.

Alan

Alan Timm

unread,
May 15, 2024, 5:27:50 PM5/15/24
to RSSC-List
Based on the limited info and how they set up the Go2 here are some guesses about the software:
  • Running on the same RockChip 3588 8G+32G board as the Go2
  • A new motion controller for the legs, probably not a full body controller
  • $16k version is "locked down" as a remote control toy
  • Pay up for the EDU version for access to the API and the bundled Jetson Orin. oh, and articulated hands
  • ROS2 CycloneDDS interfaces
  • Simple IK solvers for the arms/hands
  • A high level control API that describes robot attitude, direction, waist, and arm positions
  • Camera, Depth map and LIDAR data published
  • A low level control API using standard BLDC/SimpleFOC type protocols
  • A low speed 200Hz command and status interface
  • A high speed 500Hz command and status interface

Chris Albertson

unread,
May 15, 2024, 9:22:37 PM5/15/24
to gestalt73, RSSC-list
So it looks like the cheaper robot has plastic hands that don’t move.  This might be good because it allows the owner to experiment with hand design.   

As for software I see basic walking and self-balancing.  Notice that it occasionally takes a step to maintain balance or stepps when it is pushed.  

Also it does not seem to be able to self-balance until it is standing.  It needed a human’s hand on it’s shoulder unit we someone pressed a button on the controller to turn on the balance mode.  But this is still early.  Wait. few years and I bet the software has improved.


I wonder wht the interior is like.  Is it designed with the idea that you might want to swap out the computer for a different kind?




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

Chris Albertson

unread,
May 16, 2024, 2:16:03 AM5/16/24
to gestalt73, RSSC-list
I have been reading Unitree GitHub repository.  There is code there for controlling G1.  THere is no longer a need to speculate what the robot can do and can’t do. The API is open source and free but the on-robot code is not.  We have the definitive answer now.  See it here: https://github.com/unitreerobotics

What can you do?  Here is my quick summary….

First, we always send commands to G1 using DDS.   There is a full API that uses DDS to command the robot, so you are not using DDS directly.   The remote controller uses this interface.  In addition, ROS2 also uses DDS, so they have a wrapper you can use from ROS2. 

From Either the proprietary API or from ROS2, there are two methods for controlling the robot: “sport mode” and “motor mode” (well four methods if you multiply the modes times the APIs)

In sport mode, you have about a dozen or two commands like “stand-balance", “walk", "raise foot", "stair climb", foot position, locomotion, and hand position.  In this mode, the robot’s internal computer is handling kinematics for you.  You give commands in “robot coordinate space"

In “motor mode” you can command the position and torque of each motor and read out voltage, locations and DC current.  You would need to do your own kinematic calculations and potentially have much finer control over the robot but all you have is direct access to the motors,

Sport Mode is Ok for basic testing and demos but it can’t do things like walking on uneven surfaces and avoiding obstacles.  There is a short list of basic movements and is how you would teloperate the robot.   I think you can set the control loop to run at either 300 or 500 Hz.  (or is it 200/500?)  Sport mode programming would be relatively easy but is limited..

For more sophisticated movements you would need to run in motor mode.  Although I wonder how fast of a control loop you can do with DDS over WiFi.   Does anyone know?   is DDS fast enough to be inside a PID loop.    It seems like you’d need to run each of the 20+ motors at 500 Hz.  and read many sensors at the same rate.  (Let’s hope for WiFi 7 to be available soon.)  

So, How to do program G1 to do what I call the “hello world” of domestic robotics.  I think “fetch me a beer” is so silly and not even a well defined problem.  So I proposed to emulate Terry Winograd’s “SHRDLU”. The idea is that SHRDLU was created in 1969 by one person working alone.   If your fancy 21st-century robot can’t do what SHRDLU could do, it is not even at the “hello world” stage of development.  The test is also very well-defined.  

Here is a dialog between Winograd and his robot recorded in about 1970. https://science.slc.edu/~jmarshall/courses/cs30/Lectures/week10/SHRDLU.html

So how would you program G1 to do this?   I’d argue that if G1 can’t move colored styrofoam blocks on a tale top it is useless for any interesting task you’d expect of a humanoid robot.

As a challenge, let’s see if interested parties can’t create a “paper” describing a proposed design.  The paper should be as detailed as you typically see in refereed journals.

The Unitree GitHub has a URDF file that they say “works” in both ROS2 Rvis and Nvidia Issac environments.   Rviz would be easy to set up as it is not a simulation.  Issac would allow the authors of the above “paper” to test their proposed design without the need to spend 30K

How to control G1?  It should be clear that the controller is a computer that connects to G1 over WiFi and uses DDS. (as this is the only option Unitree provides) So the hardware can be as powerful as we want, up to a full rack of Intel Xeons or Apple Mac Studios, Nvidia H100 servers, or “whatever”.  (But there is roon for an Nvidia “orin” internally so the faster control loops need not go over WiFi.

“Problem #1”.   G1 runs in either motor mode or sport mode.  In Sport Mode, it can stand and balance. but in motor mode there is no automatic balance.  It would be nice if the legs could go into auto-balance mode while the arms run in motor mode.   This is not possible.   So it looks like we must command the robot in sport mode so we drive the robot to the table by teleoperation and then place it in “stand/balance” state.  Then command the hands to the location of the blocks as required.  The robot handles kinematics and balance.

Problem #2.  SHRDLU maintained an internal model of the blocks and remembered where each block was.   This was how all early AI worked.  Today we believe that the world is its own memory and if a robot needs to know where a block is, it looks at the block.   There is no need to remember the state of the world if you can see the world.  So G1 only needs to process one command and does not need to remember previous commands.   The overall control loop is “get command, execute command” and loops forever.   

Problem #3.   Here is where the magic happens:  Microphone —> Speech-to-text —> LLM(fine tunned for the SHRDLU problem) —> “some kind of data structure”. —> conventional robot motion planning using something like “mMoveit”.

I think it is clear that this is not going to all run on a Rasberry Pi.  Or even a Linux desktop PC.   This is a server room project. and we have not yet began to talk about wrist rotations so the fingers are aligned perpendicular to the the block surfaces.  Anyone reading out paper is going to ask about “grasp planning” and how we did that.  Then they will ask how we did force feedback on the fingers or it we used a hand design that did not need it.

The “G1 EDU” version of the robot comes with an adtional computer inside.  The Jetson Orin does not replace the 8-core ARM computer.     I think having this is a hard requirement because it solves the DDS over WiFi bottleneck by using DDS over Ethernet.  I think you want to process all the high bandwidth stuff on the robot and leave the low speed things like LLM and motion planning to the server rack.

All we need to know about what G1 can and can not do is here (no need to speculate) https://github.com/unitreerobotics





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

Alan Timm

unread,
May 16, 2024, 10:26:44 AM5/16/24
to RSSC-List
Great Find!  I never thought to check their github.


It's odd that they added it to unitree_ros.  I'm guessing that it will make it's way to the ros2 and model repositories soon.

But until then... We have URDF!

g1 isaac.jpg

Chris Albertson

unread,
May 16, 2024, 11:31:30 AM5/16/24
to gestalt73, RSSC-list
We don’t have access to the code that runs inside the robot.  So getting G1 to run in Issac means recreating that code from scratch.

On May 16, 2024, at 7:26 AM, Alan Timm <gest...@gmail.com> wrote:

Great Find!  I never thought to check their github.


It's odd that they added it to unitree_ros.  I'm guessing that it will make it's way to the ros2 and model repositories soon.

But until then... We have URDF!

<g1 isaac.jpg>

Chris Albertson

unread,
May 17, 2024, 1:10:47 AM5/17/24
to gestalt73, RSSC-list
it seems the first step is to work out both forward and inverse kinematics (homogeneous transform matrices) of the body parts relative to a “base” and then Jacobians for the same.  If the URDF is accurate (it is) then we have all the information needed to do this.

We get to define the “base link”.  If it were me making the selection I’d vote for the lower torso, below the waist as our “base”. But in the long run, it would be best to make the same choice Unitree did.  OK, I just checked.  Unitree uses the “pelvis” as the base.   This is good.

I’m reading the Unitree ROS docs and it looks like G1 is ready to go in Gazabo.  It looks like they provide launch files.  But there is a note that says the simulation accepts motor torques and you are able to move the robot using the low-level motor API.  But says the robot will not stand or walk, that is up to the user to make that happen. You can NOT use Sport Mode in the simulation.

So it looks like the first step, before you can do anything in simulation is kinematics.  We don’t actually need HD parameters, a numerical optimizer based on forward kinematics could work but you get much better performance with a closed-form IK solution and it looks possible.

Alan Timm

unread,
May 17, 2024, 11:55:22 AM5/17/24
to RSSC-List
Another view from ICRA2024

Chris Albertson

unread,
May 17, 2024, 1:55:48 PM5/17/24
to gestalt73, RSSC-list
It will take a lot of work to make G1 into a usable robot that can do real tasks.   I’m not confedent at all about runing “Motor Mode” via WiFi.  I’d want the Orin comuter insode the robot.   Of course you would still need the external server rack.

I’d really like to see photos or CAD of the interior of the robot.   Hopw much room is there inside?

Of cource the other option is to build a robot like G1.   The innovation they did to get the cost down was to make it child-size.  Remember “Mass is generally the cube of height”.  SO humanoids get heavier very quickly as they get taller.   Weight means motor torque and required power and hence batteries all grow with the cube of height

I think it would be worth exploring what it would take to make a G1-size humanoid.     You need about 24 or 28 motors and moters now costs between $80 and $250.  At $100 each you need to spend only $2800 plus the cost of some 3D printed parts for the structure, a computer, Lidar and depth camera.   You could do it for $8K, half the price of G1.

THis is been my goal for a long time.  But.  A better route is to fist build a quadruped.   They are cheaper and the software is very much the same.  Then move that technology to the more complex and expensive humanoid.

Or just skip all that and buy G1.   But G1 appears to be either a toy that does very little or a BIG software project.

But as said, the first step with G1 is kinematics.   I’m setting up, from scratch a ROS2 based  that runs in a portable container on my Mac or on a Pi4/5 or on Jetson.

Extracting HD parms from a URDF file is going to be a long and tedious problem that is very easy to get wrong.  It is worth looking for automated tools.





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

Gmail

unread,
May 17, 2024, 2:12:26 PM5/17/24
to Chris Albertson, gestalt73, RSSC-list
If it falls to the ground, it will probably break a lot of parts. 




Thomas

-
Want to learn more about ROBOTS?









On May 17, 2024, at 10:55 AM, Chris Albertson <alberts...@gmail.com> wrote:

It will take a lot of work to make G1 into a usable robot that can do real tasks.   I’m not confedent at all about runing “Motor Mode” via WiFi.  I’d want the Orin comuter insode the robot.   Of course you would still need the external server rack.

Chris Albertson

unread,
May 17, 2024, 2:31:42 PM5/17/24
to Gmail, gestalt73, RSSC-list

On May 17, 2024, at 11:12 AM, Gmail <thomas...@gmail.com> wrote:

If it falls to the ground, it will probably break a lot of parts. 

I doubt that.   This robot seems to be made of CNC aluminum parts and the rotation shafts, we know are fairly large diameter because they say the cables are routed though the hollow shafts.   All the joints would have large-diameter sealed ring bearings.

The robo-dog I’ve been working on uses 36mm outside but 24mm inside diameter ball bearing.  Two of them on every joint and the wires pass through hollow shafts.   I did not invent this, I saw it and adapted the idea to whatever bearing about that size sold at the lowest cost.  The 24x36 is $1.50 each.  So the robo-dog uses 24 of them.     It is designed to be able to tumble down a flight of stairs.   I’ve done some destructiv testing on pard made with ABS plastic.   It’s really tough and cheap.  MUCH better than my R/V servo based dog-legs.

I’m an amateur.  G1 was designed be people with more experience and it seems to re-use a lot of the technolgy from Unitree's robo-dog.  I’d expect it to survive a fall down some stairs.  This thing is NOT built with model airplane servos.  I’d worry if the sensors were protected from a fall but maybe they are resessed into the head frame?   

Also, being so short helps a lot.  Impact is the square of the height and mass is the cube of the height, being short REALLY helps.  

Even so.  damage is yet another reason to do as much software testing as possible in simulation.

Alan Timm

unread,
May 17, 2024, 2:54:37 PM5/17/24
to RSSC-List
Yeah, def not running "motor mode" over wifi.  It's all set up for an orin backpack hard wired over ethernet.  For the Go2 that's how their entire api is set up.  You have to hack it to get it to respond for high level "sportmode" over wifi.

While there hasn't been any info on the interior of the torso I can tell you that it's mostly battery.  
They give you a few hints with a render highlighting the position of the motors in the middle of the page here:
It's funny, they included the two optional torso motors even though the shells are for a 1dof waist.
I'm almost certain the optional orin board mounts somewhere inside the torso as well if equipped, it would be silly to mount it externally like a backpack.

Looking at that same render there are at least 4 different sizes of motors.  Maybe the second to smallest is the same caliber as the $100-$300 units you can find now.  It's hard for me to guess the "retail" cost of all the nonstandard motors.  That being said, you could slap together a biped i think using the $100 motors, just at a much smaller scale.

I don't think there is as much published research and code on the biped humanoid stuff, everyone's keeping their magic sauce a secret.
That being said, rumor has it to keep an eye on the Nvidia Orbit github.

Falls are still tough, the previous atlas would tuck into a ball as soon as it detected that it was falling.

Alan

Alan Timm

unread,
May 17, 2024, 2:58:05 PM5/17/24
to RSSC-List
The g1 at icra2024 took at least one spill and came out the worse for it.
If you were wondering why the left arm got stuck during initialization...

Gmail

unread,
May 17, 2024, 4:58:30 PM5/17/24
to RSSC-List
Yes, falls are tough. That’s why I will stick with wheels for my life-size 
animatronic robots for now. 




Thomas

-
Want to learn more about ROBOTS?









On May 17, 2024, at 11:58 AM, Alan Timm <gest...@gmail.com> wrote:

The g1 at icra2024 took at least one spill and came out the worse for it.
If you were wondering why the left arm got stuck during initialization...

Alan Timm

unread,
May 18, 2024, 7:29:41 PM5/18/24
to RSSC-List
Instead of doing anything productive this weekend I thought it would be fun to study the articulated hand.
Just in case there's some alternative future where they get the g1 to do useful stuff but I don't want to spend the full $40k for the version with hands.  :-)

The servos are custom and small, but it looks like it will be possible to recreate a budget version of the hand using feetech scs2332.

With servos, printing, and hardware each hand should be around $300.

unitree g1 articulated hand study.jpg



On Friday, May 17, 2024 at 1:58:30 PM UTC-7 Thomas Messerschmidt wrote:
Yes, falls are tough. That’s why I will stick with wheels for my life-size 
animatronic robots for now. 




Thomas

-
Want to learn more about ROBOTS?





alp i fall.jpg>

Alan Timm

unread,
May 18, 2024, 8:18:22 PM5/18/24
to RSSC-List
A few more clips from ICRA2024

Chris Albertson

unread,
May 18, 2024, 9:37:19 PM5/18/24
to gestalt73, RSSC-list
This is not a bad design.   It has the number one feature that hands need:  More than two fingers.  

But you can build better hands for less money.  Look at the family of hands at the “Yale Hands Project”.  Ok, I know, “better” can mean different things.  If you want to grasp object the Yale Hands win.  But is the task is to “look like a humanoid” then the Unitree hands come close to doing that.

But if the goal is to “look human” then you can do even better by using a “Brunel Hand”.   I played around with a Brunel a few years ago and discovered the Yale was better and grapging but you can’t beat Brunel for how it looks.


BTW, it is easy to experiment to grippers.   Build one and mount it to a stick and hold the stick with you left hand.   Then with right open control the hand via teleoperation.  Just make some simple/ugly sliders on a screen, whatever GUI is easy for you.    You will quickly find that crab pinchers are not very =good, three finger desinns are “ok” but the Yale “soft” or “compliant” hands are very good.  Human hands are semi-compliant.   You don’t want 100% rigidity so rubber joints beat ball bearing joints.

Now the good news.  Hands are typically built to be swapped out and you pick a (say) 30 mm bolt circle and a certain kind of electrical connector to make swapping fast.

See the girl catching the ball.   This shows what is both good and bad about human-looking hands.   They are VERY hard to control with software.   But the girl is using sensors that connect the hand to nerves in her arm.  They are the same kind of sticky sensor as if you get an ECG (electro cardiogram). SO she is controlling the bruel hand with her human brain.   If you do not have a human brain, human hands are hard to use.    So when you select a hand design think about the SOFTWARE you are going to have to write.




Yale information is here.  This hand is better functionally, Brunel is a better cosmetic fit to a humanoid.



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

Alan Timm

unread,
May 18, 2024, 11:52:10 PM5/18/24
to RSSC-List
Hey Chris,

You're not wrong.  :-)

I mostly did it because I was peeved about the $40k vs $16k penalty for working hands.
Another excuse was to take advantage of any future code that assumes the default articulated hands.

I've proven to myself that I could recreate the hand with the exact same geometry and contact areas as the original.
The only functional difference would be that you can no longer hold things between the two fingers like in the original.

Now that I've stared at them for a day I'm starting to wonder what they intended for these hands to do.
They can't open doors.  They can hold bottles glasses and other things.  Maybe they could hold a plate.

There are already quite a few open source, commercial, and almost ready for sale hands out there right now.
The next 12 months are going to be really interesting.  :-)


budget hand study.jpg

Alan Timm

unread,
May 19, 2024, 4:18:40 PM5/19/24
to RSSC-List
After modeling some of the required brackets I see it's time for a plan b.  mitten hands will never do.
The thumb tip must fit in the space between the two fingers.  

Time to get creative.

mitten hands.jpg

On Saturday, May 18, 2024 at 8:52:10 PM UTC-7 Alan Timm wrote:
Hey Chris,

You're not wrong.  :-)

I mostly did it because I was peeved about the $40k vs $16k penalty for working hands.
Another excuse was to take advantage of any future code that assumes the default articulated hands.

I've proven to myself that I could recreate the hand with the exact same geometry and contact areas as the original.
The only functional difference would be that you can no longer hold things between the two fingers like in the original.

Now that I've stared at them for a day I'm starting to wonder what they intended for these hands to do.
They can't open doors.  They can hold bottles glasses and other things.  Maybe they could hold a plate.

There are already quite a few open source, commercial, and almost ready for sale hands out there right now.
The next 12 months are going to be really interesting.  :-)



Chris Albertson

unread,
May 19, 2024, 4:36:16 PM5/19/24
to gestalt73, RSSC-list
One OTHER show-stopper with your design.  You seem to nbe using R/C servoes.  These are commanded to a position angle by PWM.   Fingers should be controlled by torque.  Why?  Because it is not enough to close you fingers around an object, you need to also have just the right amount of pressure so as to create just enough friction.    How much friction and pressure?  Try this experiment:  Pick up a full beer can and see how tightly you need to hold it.    Then try this with an empty can, you crush it if you applied that much pressure.   What wait?  the can is EXACTLY the same size and shape.   How can an R/C servo based hand move the finger to the same exact place but also control the pressure?

The answer is compliance.   The hand you design has zero compliance.  it is 100% full-on ridgid and would not work well.

The engineering reason for the need for compliance is “Hooke’s Law”. A pring will convert position to force.   The servoes must act on the finfer via some kind of spring.   You can simply mount the servos in rubber and make the finger of rubber.  But some how you need lots of “springly stuff” in the drive chain or in the structure.

Geting thios right is just rail and error.   So I said to build a known-to-work hand.  The Brunel hand uses strings and rubber hinges to emulate a human hand.  The Yale hands also use strings and rubber but gave up on even trying to look humanoid.

Also as I found on my Quadrupeds,  “moment matters”.  You really, really want the moters to be as close to the body as possible.   What matrters is “mass times the distance”.   Why, because with less moment you can use a smaller wrist motor which means a smaller elbow motor which means a smaller shoulder motor…. It all compounds and every “gam centimeter” counts.    My next quad will need to made all the motors pulled back and the mommentsoin the current design are not good.

In any case, no matter how it is done, rigid directly driven fingers don’t work.  After that is is tjust details.  You need to be able to control the force on the fingertip even if they don’t move.

One more hind.  BNefore designing the hand, look at some grasp-planning software.  See what the software wants.  


On May 19, 2024, at 1:18 PM, Alan Timm <gest...@gmail.com> wrote:

After modeling some of the required brackets I see it's time for a plan b.  mitten hands will never do.
The thumb tip must fit in the space between the two fingers.  

Time to get creative.
-- 
You received this message because you are subscribed to the Google Groups "RSSC-List" group.
To unsubscribe from this group and stop receiving emails from it, send an email to rssc-list+...@googlegroups.com.

Alan Timm

unread,
May 19, 2024, 6:40:22 PM5/19/24
to RSSC-List
Oh I'm designing with Feetech SCS2332 Serial Bus Servos.
I had a few lying around from a previous project.


There is a standardish pwm servo (K989-58) that would fit in the finger envelope, but I don't want to design a servo hand without feedback (position, temp, load etc)

I still haven't quite figured out how they demonstrated force compliance in the video.  If I had to guess they're monitoring load and disengaging the servo at overload condition.
scs2332.jpg

Alan Timm

unread,
May 22, 2024, 11:09:12 PM5/22/24
to RSSC-List
So I ended up modeling the most of the hand using the urdf file, photos, and video as reference.

It took me a minute to figure out why the thumb pivots. I's so a finger and thumb can pinch together.
there's a bunch of weird angles on the fingertip but if the thumb is at a specific angle the two finger tips should connect at the flat part.

A few folks have been given the pricing sheet, and only the $55k version has the articulated hands.  That version also has 1+2 dof wrists and the 1+2dof waist.

They offer to sell the hands separately @ $5000.  for. each. hand.


unitree g1 hand pinch.jpg

Thomas Messerschmidt

unread,
May 23, 2024, 12:11:21 AM5/23/24
to Alan Timm, RSSC-List
That's a really great model Alan! How long did it take? Are you going to print a working version of it? 

--
You received this message because you are subscribed to the Google Groups "RSSC-List" group.
To unsubscribe from this group and stop receiving emails from it, send an email to rssc-list+...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/rssc-list/ae68a662-4528-411b-bb34-1584a0bb28fdn%40googlegroups.com.

Chris Albertson

unread,
May 23, 2024, 12:39:44 AM5/23/24
to gestalt73, RSSC-list
This is looking good so far.   I think with some effort this 3D model could be made into something that could 3D printed.

I looked at the fusion model.  Yes, it looks like it was taken from the URDF file.   I think these are not parts you would manufacture.  They are designed for renderings.  Also, the motors do not look like standard R/C servos.  But all this can be fixed.  So long as the joints and finger surfaces are faithful to the URDF.

What’s next? The arm and shoulder?   Given that they have published a URDF it should be possible to model the entire robot and even print a copy

A printed G1 clone would cost quite a bit less than $16K but it would also not have the strength or power of a real G1

About the price of the hands.   Those motors are likely not RC servos and might cost $100 or more each.  They might be BLDC gear motors with controllable torque.  Likely more than $100 each. and then there are the fingerpad pressure sensors


As for the waist and wrists.   I think you need those if you are going to put the hands to use.   Obviously, the wrists are needed so the hands can have the angle adjusted for grasping.   The waist is needed for balance while holding a weight with the hand(s).    The robot would need the ability to lean back if it holds something. It could lead from the ankles but then the leg angles are not good for walking.




On May 22, 2024, at 8:09 PM, Alan Timm <gest...@gmail.com> wrote:

So I ended up modeling the most of the hand using the urdf file, photos, and video as reference.

It took me a minute to figure out why the thumb pivots. I's so a finger and thumb can pinch together.
there's a bunch of weird angles on the fingertip but if the thumb is at a specific angle the two finger tips should connect at the flat part.

A few folks have been given the pricing sheet, and only the $55k version has the articulated hands.  That version also has 1+2 dof wrists and the 1+2dof waist.

They offer to sell the hands separately @ $5000.  for. each. hand.


<unitree g1 hand pinch.jpg>



On Sunday, May 19, 2024 at 3:40:22 PM UTC-7 Alan Timm wrote:
Oh I'm designing with Feetech SCS2332 Serial Bus Servos.
I had a few lying around from a previous project.


There is a standardish pwm servo (K989-58) that would fit in the finger envelope, but I don't want to design a servo hand without feedback (position, temp, load etc)

I still haven't quite figured out how they demonstrated force compliance in the video.  If I had to guess they're monitoring load and disengaging the servo at overload condition.


-- 
You received this message because you are subscribed to the Google Groups "RSSC-List" group.
To unsubscribe from this group and stop receiving emails from it, send an email to rssc-list+...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/rssc-list/ae68a662-4528-411b-bb34-1584a0bb28fdn%40googlegroups.com.
<unitree g1 hand pinch.jpg>

Alan Timm

unread,
May 23, 2024, 5:19:21 PM5/23/24
to RSSC-List
Thomas:  Thanks!  I've been working on it on and off since last weekend.  Without serial bus servos that fit in the fingers there's not much point. It would be possible to print out a 1:1 model though.

Chris:  Yeah, I used the urdf stls for dimensional reference, then pics and screenshots from their promotional materials to recreate the details.  Every part for the hand is cnc machined in the real version except maybe the palm which maybe is cast.  If I were to guess they're coreless motor serial bus servos with beefy gears, and they're very custom.  Also there's no separate left and right hand design.  It's the same hand flipped 180 degrees.

g1 hand reference 1.jpg

Chris Albertson

unread,
May 23, 2024, 5:37:08 PM5/23/24
to gestalt73, RSSC-list
You are right about the motors.  They are likely custom-made.  But if you really did want to make a hand like this. You could modify the design.  Replace the motors with a rubber spring that extends the fingers.   Then you close them with a kevlar cord and one servo per finger and two for the thumb so it can pivot.   Many very good hands work using string “tendons”.

Both the Yale and Brunel hands that I pointed put use tendons made of no-stretch cord that connect to servos.  Human hands are like this too.





On May 23, 2024, at 2:19 PM, Alan Timm <gest...@gmail.com> wrote:

Thomas:  Thanks!  I've been working on it on and off since last weekend.  Without serial bus servos that fit in the fingers there's not much point. It would be possible to print out a 1:1 model though.

Chris:  Yeah, I used the urdf stls for dimensional reference, then pics and screenshots from their promotional materials to recreate the details.  Every part for the hand is cnc machined in the real version except maybe the palm which maybe is cast.  If I were to guess they're coreless motor serial bus servos with beefy gears, and they're very custom.  Also there's no separate left and right hand design.  It's the same hand flipped 180 degrees.

<g1 hand reference 1.jpg>



Alan Timm

unread,
May 24, 2024, 10:14:43 AM5/24/24
to RSSC-List
Here are two more clips from ICRA2024.

Robot Parade at ICRA2024, unitree g1 visible at 57 seconds

Unitree G1, Humanoid AI avatar (ICRA 2024), first view of unitree g1's back and rear mechanics at 13 seconds

Alan Timm

unread,
May 24, 2024, 11:35:36 AM5/24/24
to RSSC-List
Man things are moving quickly.

Unitree G1 reinforcement learning for locomotion is already supported in loco mujoco:
https://github.com/robfiras/loco-mujoco
https://loco-mujoco.readthedocs.io/en/latest/

https://www.youtube.com/watch?v=t2mnIxpisxQ
loco.jpg



On Friday, May 17, 2024 at 1:58:30 PM UTC-7 Thomas Messerschmidt wrote:

Chris Albertson

unread,
May 24, 2024, 8:15:02 PM5/24/24
to gestalt73, RSSC-list

On May 24, 2024, at 8:35 AM, Alan Timm <gest...@gmail.com> wrote:

Man things are moving quickly.

Yes, the files to enable this were released just last week — https://github.com/unitreerobotics/unitree_ros/tree/master/robots/g1_description

The first G1 will be delivered to customers in about 6 months.  We will know a lot more after “real people” have some experience with them







Unitree G1 reinforcement learning for locomotion is already supported in loco mujoco:
https://github.com/robfiras/loco-mujoco
https://loco-mujoco.readthedocs.io/en/latest/

https://www.youtube.com/watch?v=t2mnIxpisxQ
<loco.jpg>



On Friday, May 17, 2024 at 1:58:30 PM UTC-7 Thomas Messerschmidt wrote:
Yes, falls are tough. That’s why I will stick with wheels for my life-size 
animatronic robots for now. 


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

Alan Timm

unread,
May 25, 2024, 5:17:50 PM5/25/24
to RSSC-List
Instead of doing anything productive this weekend I thought I'd take another crack at the hand design.

I found another option that maintains the geometry and general aesthetics of the original by using commonly available bevel gears.
Something like these:  https://www.aliexpress.us/item/3256806372282470.html

I'm hoping to add in finger splay [abduction/adduction] as well.  I'll see how much room I have left after I update the design.

unitree g1 finger drive.jpg

Alan Timm

unread,
May 28, 2024, 4:06:18 PM5/28/24
to RSSC-List
a few more video clips showing the g1 as well as it's capabilities/limitations:

A few things become obvious after seeing it outside of the curated promotional video. Notice how everyone holds it steady by the neck?  reliable walking just isn't possible without RL, and at the current state of the software running on the rockchip 3588 it can't be trusted to walk without assistance.

In other words there's nothing to get excited about yet until you pair up the hardware with a beefy jetson orin and a few software breakthroughs.  What do you think?  6 months until it's useful?  24 months?

Also the 9000mah battery was throwing me off.  2 hours of operating time is short.  From the outside it looks like a standard go2 battery.  I was confused about why they wouldn't just use the 15000mah go2 battery and why the mah didn't match the standard go2 battery capacity. I think they're using the high capacity cells in a version of the pack that's only 2/3 as deep, and that all the electronics are fit between the battery and the far side of the shell.  They can't use any of the room below the battery because of the plans for the 3dof waist option, and all the space above is probably taken up by the shoulder motors.

g1 with short battery stack.jpg

Chris Albertson

unread,
May 28, 2024, 5:27:17 PM5/28/24
to gestalt73, RSSC-list
Why are you buying this?

  1. if you want a robot that can clean your bathroom and then change the oil in your car you will wait a few decades and it will not be as cheap as $16K.
  2. If you want some pretty good hardware so you can experiment with different software control ideas you have then buy it NOW.  If you want this but don’t have $16K then put it inside a simulator NOW.
  3. Something different? Maybe a more capable toy robot?     Then it comes out of your entertainment budget.

I think Unitree is building these for Cases #2 and #3.   It is either for research or a toy.


I really doubt the limitation is the rock chip computer.   I know first hand that a less-capable Pi4 can do about 2,000 kinemetic calculations per second using just one of the four cores if written in Python and running Ubuntu.  That is more then enough for dynamic balance.

Watch what the handler is doing when he places his hand on the robot.   It seems tobe either (1) The balancemode command has not yet been issued from the remote controller or (2) is controllering the direction the robot walks by controlling it’s balance.

Turns are hard. I am not sure I have the mathright even for 3-DOF legs with sherical feet.  It would be much harder for a 5-DOF biped

Then watch the boxing glove demo.  The robot does a good job of statying on its feet

What this robot can’t do is any kind of purposfull behavior.  there is not one bit of “imtelegence.”   It apears not even the very simple ROS-based SLAM navigation.   In fact we have not yet seen this robot walk around a corner of make a turn, only forward and backward walking.   But this is perfectly good enough if the purpose for buying it is #2

Based on other Unitree products don’t wait for this to be a usable domestic robot.  That is not what they do.  You would either have to DIY programming or hire a consultant. Even Boston Dynamic is selling only “stock” robots that do basically nothing untill you either program them or hire someone to program them.




On May 28, 2024, at 1:06 PM, Alan Timm <gest...@gmail.com> wrote:

a few more video clips showing the g1 as well as it's capabilities/limitations:

A few things become obvious after seeing it outside of the curated promotional video. Notice how everyone holds it steady by the neck?  reliable walking just isn't possible without RL, and at the current state of the software running on the rockchip 3588 it can't be trusted to walk without assistance.

In other words there's nothing to get excited about yet until you pair up the hardware with a beefy jetson orin and a few software breakthroughs.  What do you think?  6 months until it's useful?  24 months?

Also the 9000mah battery was throwing me off.  2 hours of operating time is short.  From the outside it looks like a standard go2 battery.  I was confused about why they wouldn't just use the 15000mah go2 battery and why the mah didn't match the standard go2 battery capacity. I think they're using the high capacity cells in a version of the pack that's only 2/3 as deep, and that all the electronics are fit between the battery and the far side of the shell.  They can't use any of the room below the battery because of the plans for the 3dof waist option, and all the space above is probably taken up by the shoulder motors.

<g1 with short battery stack.jpg>


On Friday, May 24, 2024 at 8:35:36 AM UTC-7 Alan Timm wrote:
Man things are moving quickly.

Unitree G1 reinforcement learning for locomotion is already supported in loco mujoco:
https://github.com/robfiras/loco-mujoco
https://loco-mujoco.readthedocs.io/en/latest/

https://www.youtube.com/watch?v=t2mnIxpisxQ
loco.jpg



On Friday, May 17, 2024 at 1:58:30 PM UTC-7 Thomas Messerschmidt wrote:
Yes, falls are tough. That’s why I will stick with wheels for my life-size 
animatronic robots for now. 


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

Gmail

unread,
May 28, 2024, 6:10:19 PM5/28/24
to Alan Timm, RSSC-List
It is being steered by its neck!




Thomas

-
Want to learn more about ROBOTS?









On May 28, 2024, at 1:06 PM, Alan Timm <gest...@gmail.com> wrote:

a few more video clips showing the g1 as well as it's capabilities/limitations:

A few things become obvious after seeing it outside of the curated promotional video. Notice how everyone holds it steady by the neck?  reliable walking just isn't possible without RL, and at the current state of the software running on the rockchip 3588 it can't be trusted to walk without assistance.

In other words there's nothing to get excited about yet until you pair up the hardware with a beefy jetson orin and a few software breakthroughs.  What do you think?  6 months until it's useful?  24 months?

Also the 9000mah battery was throwing me off.  2 hours of operating time is short.  From the outside it looks like a standard go2 battery.  I was confused about why they wouldn't just use the 15000mah go2 battery and why the mah didn't match the standard go2 battery capacity. I think they're using the high capacity cells in a version of the pack that's only 2/3 as deep, and that all the electronics are fit between the battery and the far side of the shell.  They can't use any of the room below the battery because of the plans for the 3dof waist option, and all the space above is probably taken up by the shoulder motors.

<g1 with short battery stack.jpg>



On Friday, May 24, 2024 at 8:35:36 AM UTC-7 Alan Timm wrote:
Man things are moving quickly.

Unitree G1 reinforcement learning for locomotion is already supported in loco mujoco:
https://github.com/robfiras/loco-mujoco
https://loco-mujoco.readthedocs.io/en/latest/

https://www.youtube.com/watch?v=t2mnIxpisxQ
loco.jpg



On Friday, May 17, 2024 at 1:58:30 PM UTC-7 Thomas Messerschmidt wrote:
Yes, falls are tough. That’s why I will stick with wheels for my life-size 
animatronic robots for now. 

--
You received this message because you are subscribed to the Google Groups "RSSC-List" group.
To unsubscribe from this group and stop receiving emails from it, send an email to rssc-list+...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/rssc-list/8deb8e66-88cd-49a8-b6d3-82ff39837eban%40googlegroups.com.

Chris Albertson

unread,
May 28, 2024, 6:38:01 PM5/28/24
to Gmail, gestalt73, RSSC-list

On May 28, 2024, at 3:09 PM, Gmail <thomas...@gmail.com> wrote:

It is being steered by its neck!

Yes, and the fact that this is possible says a lot about how good the balance is.  If the balance were poor the robot would fall down when pushed

Alan Timm

unread,
Aug 19, 2024, 10:04:18 AM8/19/24
to RSSC-List
A new Unitree video showing the capabilities of the hardware.  I think they switched to an rl walking model which would explain the abuses it tolerates now.  It also walks up and down stairs, but without a downward facing depth camera I'm not sure how they're doing it.  And it walked into a go2 towards the end.  :-/

"Over the past few months, Unitree G1 robot has been upgraded into a mass production version, with stronger performance, ultimate(final) appearance, and being more in line with mass production requirements. We hope you like it.🥳"

g1 jumpy jumpy.jpg

Alan Timm

unread,
Aug 27, 2024, 12:07:01 PM8/27/24
to RSSC-List
Another quick video that gives a a really nice close up view of what they're calling final hardware:

They have the hang loops on top and a handle on the back now.
This one also teases the 3dof waist.  I like the look of the waist better but don't see a reason for the extra 2dof in this application.

Chris Albertson

unread,
Aug 27, 2024, 1:24:35 PM8/27/24
to gestalt73, RSSC-list

On Aug 27, 2024, at 9:07 AM, Alan Timm <gest...@gmail.com> wrote:

Another quick video that gives a a really nice close up view of what they're calling final hardware:

They have the hang loops on top and a handle on the back now.
This one also teases the 3dof waist.  I like the look of the waist better but don't see a reason for the extra 2dof in this application.

The 3DOF waist is going to be very important if you want to place a payload in the hands.   Yes for this application with dummy/nonfunctional hands, you do not need it.   

Think abut the task where the robot needs to lift an object off of a table.   The robot extends the hand, grips an object and then lifts it upward,  It now has some mass in front of it.    It could remain balanced if it leaned back from the ankles but that would place the entire robot torso in a backward off-vertical position.   It would be best to lean back at the ankles but also forward at the waist.

Leaning back only with ankles means the hands have to move forward to maintain the grip on the object being lifted.   The better wait allows the shoulders to remain the same distance from the table.

The big problem is lifting objects.    While doing so the center of gravity changes as the hands apply lifting force but usually the feet and hands are in fixed position untill enough lifting force is applied that the oject is liffted off the table and is in free space,   Only after that can the hands move.   So some way is needed to maintain balance.


Next, you do need all those DOF if the robot is not square to the table as leaning forward is now a 6-dof movement, yes “6” because you want to control the position and orientation of the shoulders.

The next reason for a 3 or 4 DOF waits is dynamic walking.   It seems the Unitree robot is now only able to do a kind of “basic shuffle” walk.  I have not seen even large human-like walking strides let alone true running where both feet leave the ground at the same time. (both feet off the ground is the definition of “running”)

I think the basic human-like robot needs 28-DOF, not counting the hands.

But on the other end of the use-case spectrum, if the use case is a social robot that acts as the physical embodiment of an LLM “chatbot”  then only very basic mobility is needed



On Monday, August 19, 2024 at 7:04:18 AM UTC-7 Alan Timm wrote:
A new Unitree video showing the capabilities of the hardware.  I think they switched to an rl walking model which would explain the abuses it tolerates now.  It also walks up and down stairs, but without a downward facing depth camera I'm not sure how they're doing it.  And it walked into a go2 towards the end.  :-/

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

Gmail

unread,
Aug 27, 2024, 2:09:05 PM8/27/24
to Alan Timm, RSSC-List
Alan,

It certainly would run circles around your Inmoov robot. Excellent balance. 

So are you actually going to buy this thing? 



Thomas

-  

Need something prototyped, built or coded? I’ve been building prototypes for companies for 15 years. I am now incorporating generative AI into products.

-

Need a great hardworking engineer? I am currently looking for a new job opportunity in robotics and/ or AI. 

Contact me directly or through LinkedIn:   


On Aug 27, 2024, at 9:07 AM, Alan Timm <gest...@gmail.com> wrote:

Another quick video that gives a a really nice close up view of what they're calling final hardware:
--

Alan Timm

unread,
Sep 13, 2024, 2:22:40 PM9/13/24
to RSSC-List
Here's another glimpse of what they're calling "final hardware":

This configuration would match what they're calling "Unitree G1 Edu Smart" with 7dof arms and a 3dof waist.
Here you can see their more standard approach to a 2dof wrist and how the 3dof waist looks different than their new 1dof waist.

(This also suggests that their base model will also have an exposed midriff, which I think looks better anyways.)

unitree g1 edu smart .jpg

Gmail

unread,
Sep 13, 2024, 3:28:42 PM9/13/24
to Alan Timm, RSSC-List
Do you have another link for "final hardware"? The link was removed. 🙁




Thomas

-  

Need something prototyped, built or coded? I’ve been building prototypes for companies for 15 years. I am now incorporating generative AI into products.

-

Need a great hardworking engineer? I am currently looking for a new job opportunity in robotics and/ or AI. 

Contact me directly or through LinkedIn:   


On Sep 13, 2024, at 11:22 AM, Alan Timm <gest...@gmail.com> wrote:

Here's another glimpse of what they're calling "final hardware":

This configuration would match what they're calling "Unitree G1 Edu Smart" with 7dof arms and a 3dof waist.
Here you can see their more standard approach to a 2dof wrist and how the 3dof waist looks different than their new 1dof waist.

(This also suggests that their base model will also have an exposed midriff, which I think looks better anyways.)

<unitree g1 edu smart .jpg>




On Tuesday, August 27, 2024 at 11:09:05 AM UTC-7 Thomas Messerschmidt wrote:
Alan,

It certainly would run circles around your Inmoov robot. Excellent balance. 

So are you actually going to buy this thing? 



Thomas

-  

Need something prototyped, built or coded? I’ve been building prototypes for companies for 15 years. I am now incorporating generative AI into products.

-

Need a great hardworking engineer? I am currently looking for a new job opportunity in robotics and/ or AI. 

Contact me directly or through LinkedIn:   

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

Alan Timm

unread,
Sep 13, 2024, 3:44:11 PM9/13/24
to RSSC-List
Huh, that's weird...  Here's another link.  (X... amirite?)

The Unitree G1 Edu Smart will be $45,000 according to their pricing sheet.

If you want that, plus motorized hands and fingers the Unitree G1 Edu Ultimate will set you back $55,000.

Too rich for my blood, but hopefully the base $16,000 version will be hackable.  fingers crossed!

Alan

Gmail

unread,
Sep 13, 2024, 4:33:26 PM9/13/24
to Alan Timm, RSSC-List
That works. Thanks. Yes $55,000 is much too much money. Have you  already ordered one of the $16,000 robots? 




Thomas

-  

Need something prototyped, built or coded? I’ve been building prototypes for companies for 15 years. I am now incorporating generative AI into products.

-

Need a great hardworking engineer? I am currently looking for a new job opportunity in robotics and/ or AI. 

Contact me directly or through LinkedIn:   


On Sep 13, 2024, at 12:44 PM, Alan Timm <gest...@gmail.com> wrote:

Huh, that's weird...  Here's another link.  (X... amirite?)
--
You received this message because you are subscribed to the Google Groups "RSSC-List" group.
To unsubscribe from this group and stop receiving emails from it, send an email to rssc-list+...@googlegroups.com.

Alan Timm

unread,
Oct 17, 2024, 11:25:29 AM10/17/24
to RSSC-List
Another video update on the Unitree G1 showing off the hardware capabilities.

And another video from a tradeshow demonstrating why it might not yet be ready for consumer release:

In this second video it's easy to see what went wrong:
  • different behaviors are triggered and governed by different scripts
  • the jump behavior was coded/trained on hard flat ground
  • there's no recovery in the jump behavior
  • the added springiness of the carpet caused a slight tilt back in the jump
  • after the jump script concludes the standing / walking model takes over
  • the standing / walking model always assumes that you're upright
  • commence with the flailing

These should be hitting the market soon, I can't wait!


2024-10-17 g1 plyojump.jpg

On Friday, September 13, 2024 at 1:33:26 PM UTC-7 Thomas Messerschmidt wrote:
That works. Thanks. Yes $55,000 is much too much money. Have you  already ordered one of the $16,000 robots? 

Thomas

-  Need something prototyped, built or coded? I’ve been building prototypes for companies for 15 years. I am now incorporating generative AI into products.

-  Need a great hardworking engineer? I am currently looking for a new job opportunity in robotics and/ or AI. 

Contact me directly or through LinkedIn:   

James Mobile

unread,
Oct 18, 2024, 2:37:10 PM10/18/24
to Alan Timm, RSSC-List
Tuesday I got a marketing email from the Robot Shop indicating that they are now carrying the Unitree line. Looked like the complete line, but I didn’t check it that completely. 

On Oct 17, 2024, at 08:25, Alan Timm <gest...@gmail.com> wrote:

Another video update on the Unitree G1 showing off the hardware capabilities.

And another video from a tradeshow demonstrating why it might not yet be ready for consumer release:

In this second video it's easy to see what went wrong:
  • different behaviors are triggered and governed by different scripts
  • the jump behavior was coded/trained on hard flat ground
  • there's no recovery in the jump behavior
  • the added springiness of the carpet caused a slight tilt back in the jump
  • after the jump script concludes the standing / walking model takes over
  • the standing / walking model always assumes that you're upright
  • commence with the flailing

These should be hitting the market soon, I can't wait!


<2024-10-17 g1 plyojump.jpg>


On Friday, September 13, 2024 at 1:33:26 PM UTC-7 Thomas Messerschmidt wrote:
That works. Thanks. Yes $55,000 is much too much money. Have you  already ordered one of the $16,000 robots? 

Thomas

-  Need something prototyped, built or coded? I’ve been building prototypes for companies for 15 years. I am now incorporating generative AI into products.

-  Need a great hardworking engineer? I am currently looking for a new job opportunity in robotics and/ or AI. 

Contact me directly or through LinkedIn:   

--
You received this message because you are subscribed to the Google Groups "RSSC-List" group.
To unsubscribe from this group and stop receiving emails from it, send an email to rssc-list+...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/rssc-list/d3c25c85-6605-4c31-a807-6bb715cc3b61n%40googlegroups.com.
<2024-10-17 g1 plyojump.jpg>

Alan Timm

unread,
Dec 21, 2024, 1:06:49 PM12/21/24
to RSSC-List
Looks like G1s are being shipped.

"if this is what it took to jailbreak C-3PO no wonder Anakin turned to the dark side"
https://x.com/robertghilduta/status/1870005662299893847

he has a few more posts about the g1 on his Xitter feed.

Chris Albertson

unread,
Dec 21, 2024, 3:08:20 PM12/21/24
to gestalt73, RSSC-list
Most of us understand that this is just a frame with some motors attached and the little humaniod is dumber than a toaster oven.

But most of the public is different.   They see a human shape and expect human behavior.     For example by daughter, she has a decent education.  She is a graduate student with a sciene (not engineering) degree.  But she saw this and thought “I’d buy one in a minute if they are under $20K.”  She sees it walking and thought that if it can walk then it only needs some “simple software” and it shoud be able to do laundry and house-keeping and maybe cooking, painting, and gardening.   A slave that can work 24x7 for years for only $20K, it’s a no-brainer.

How long before I can have a robot where I can say “Hey Robbie, the tenents moved out of 195th St. Unit E, I want you to clean it for a move in but first there is some painting and I need you to replace the trim on that door where it is messed-up and change ther HVAC filter too.  Go over there and look, and tell me how long it will take for you to make the unit ready for a move-in.   Then Robbie makes a plan, tells me the extimated completion date and starts working.   Robbie is humanoid so he fits inside a taxi and can transports himself to the apartment building.  I can give Robbie my credit card and he can order materials and suplies to be delived to the job site.

How long until this “simple software” is ready?




On Dec 21, 2024, at 10:06 AM, Alan Timm <gest...@gmail.com> wrote:

Looks like G1s are being shipped.

"if this is what it took to jailbreak C-3PO no wonder Anakin turned to the dark side"
https://x.com/robertghilduta/status/1870005662299893847

he has a few more posts about the g1 on his Xitter feed.
<hacking a g1.jpeg>


On Friday, October 18, 2024 at 11:37:10 AM UTC-7 James Bond wrote:
Tuesday I got a marketing email from the Robot Shop indicating that they are now carrying the Unitree line. Looked like the complete line, but I didn’t check it that completely. 
--
You received this message because you are subscribed to the Google Groups "RSSC-List" group.
To unsubscribe from this group and stop receiving emails from it, send an email to rssc-list+...@googlegroups.com.

Alan Timm

unread,
Dec 21, 2024, 3:29:48 PM12/21/24
to RSSC-List
I think we're going to get there faster than you expect.  And this "toaster-level-intelligence" hardware will be able to take advantage of it when we do.

My Alfie is also waiting for that small handful of software breakthroughs.  My personal goalpost is anything as capable as a 6-8 year old.
"Honey, can you take this and hand it to dad in the kitchen" type of thing.  or put it in the sink.  or throw this in the trash.  start small, dream big.  :-)

Another fanciful look at the near future (turn on english captions):

g1 pretending to do the dishes.jpg

Gmail

unread,
Dec 21, 2024, 3:39:19 PM12/21/24
to Alan Timm, RSSC-List
I just want all of my appliances to have “real people personalities” along the lines of the HHGTTG. 

(Ghastly!)




Thomas

-  

Need something prototyped, built or coded? I’ve been building prototypes for companies for 15 years. I am now incorporating generative AI into products.

-

Need a great hardworking engineer? I am currently looking for a new job opportunity in robotics and/ or AI. 

Contact me directly or through LinkedIn:   


On Dec 21, 2024, at 3:29 PM, Alan Timm <gest...@gmail.com> wrote:

I think we're going to get there faster than you expect.  And this "toaster-level-intelligence" hardware will be able to take advantage of it when we do.

My Alfie is also waiting for that small handful of software breakthroughs.  My personal goalpost is anything as capable as a 6-8 year old.
"Honey, can you take this and hand it to dad in the kitchen" type of thing.  or put it in the sink.  or throw this in the trash.  start small, dream big.  :-)

Another fanciful look at the near future (turn on english captions):

<g1 pretending to do the dishes.jpg>


On Saturday, December 21, 2024 at 12:08:20 PM UTC-8 albertson.chris wrote:
Most of us understand that this is just a frame with some motors attached and the little humaniod is dumber than a toaster oven.

But most of the public is different.   They see a human shape and expect human behavior.     For example by daughter, she has a decent education.  She is a graduate student with a sciene (not engineering) degree.  But she saw this and thought “I’d buy one in a minute if they are under $20K.”  She sees it walking and thought that if it can walk then it only needs some “simple software” and it shoud be able to do laundry and house-keeping and maybe cooking, painting, and gardening.   A slave that can work 24x7 for years for only $20K, it’s a no-brainer.

How long before I can have a robot where I can say “Hey Robbie, the tenents moved out of 195th St. Unit E, I want you to clean it for a move in but first there is some painting and I need you to replace the trim on that door where it is messed-up and change ther HVAC filter too.  Go over there and look, and tell me how long it will take for you to make the unit ready for a move-in.   Then Robbie makes a plan, tells me the extimated completion date and starts working.   Robbie is humanoid so he fits inside a taxi and can transports himself to the apartment building.  I can give Robbie my credit card and he can order materials and suplies to be delived to the job site.

How long until this “simple software” is ready?




On Dec 21, 2024, at 10:06 AM, Alan Timm <gest...@gmail.com> wrote:

Looks like G1s are being shipped.

"if this is what it took to jailbreak C-3PO no wonder Anakin turned to the dark side"
https://x.com/robertghilduta/status/1870005662299893847

he has a few more posts about the g1 on his Xitter feed.
<hacking a g1.jpeg>


On Friday, October 18, 2024 at 11:37:10 AM UTC-7 James Bond wrote:
Tuesday I got a marketing email from the Robot Shop indicating that they are now carrying the Unitree line. Looked like the complete line, but I didn’t check it that completely. 
--
You received this message because you are subscribed to the Google Groups "RSSC-List" group.
To unsubscribe from this group and stop receiving emails from it, send an email to rssc-list+...@googlegroups.com.

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

Chris Albertson

unread,
Dec 21, 2024, 6:24:43 PM12/21/24
to gestalt73, RSSC-list
I the video, she is talking like I’m, about what I woulld like to see in the future.     Notice the tricks 
(1) this is the low-cost version with non-fuctional hands, ther cup is glued to the hand because the fingrs don’t move. And
(2) the closeup of the hand is a human wearing a black glove.

I’m not faulting the video as she is saying this si what the future could be like.

What is needed is a way to get from an LLM’s language abilty to the low level controls in a robot.     I think it can be done and I’m starting with somethis a LOT less complex and expensive.  I’m usinf a light switch.   The Home Assistant people just released and open source competitor to Sir that can run locally.


On Dec 21, 2024, at 12:29 PM, Alan Timm <gest...@gmail.com> wrote:

I think we're going to get there faster than you expect.  And this "toaster-level-intelligence" hardware will be able to take advantage of it when we do.

My Alfie is also waiting for that small handful of software breakthroughs.  My personal goalpost is anything as capable as a 6-8 year old.
"Honey, can you take this and hand it to dad in the kitchen" type of thing.  or put it in the sink.  or throw this in the trash.  start small, dream big.  :-)

Another fanciful look at the near future (turn on english captions):

<g1 pretending to do the dishes.jpg>

Alan Timm

unread,
Jan 12, 2025, 1:41:40 PMJan 12
to RSSC-List
https://www.youtube.com/watch?v=yYh4yw6fbUY

Another look at the G1 from CES 2025 thanks to KhanFlicks.

A few observations:
  • the final design has the exposed midriff.  probably for cost savings but I think it looks better
  • another video mentioned that the $16k version was china only, not confirmed
  • the back bar and shoulder hooks appear to be standard
  • The head has a pivot up and down with a manual joint in the neck
  • The sheen on most of the body arm and leg panels have changed. not sure if they're still aluminum or if the final version is plastic
  • we're starting to see these in the wild

Also, The G1 and recent Go2s have had SecureBoot enabled, so looks like the party's over for 3rd party hacking for now.  boo.

We have a guy on the hacking forum that bought two G1s and accidentally bricked one in an attempt to hack the sbc.  :-(

Really hard to justify a $16k remote controlled robot that isn't programmable.  Aside from that everyone's gearing up for that handful of software breakthroughs that will make these things useful.  I'd stay tuned.

Alan

Chris Albertson

unread,
Jan 12, 2025, 5:45:49 PMJan 12
to gestalt73, RSSC-list
Yes, $16K for a closed sourse remote control toy with no usable hands.   No hands is a big deal. Not programable is a big deal too.

What’s needed is a good open source humanoid project.   I think you could get the cost of entry way down if it were possible to build a partial robot, say a fixed stand with one arm and head-mounted sensors.  It could do many table top tasks.   Building out all four limbs and a torso would cost more, but could be done in stages.

A full humanoid has about 26 degrees of freedom if you do not count the hands.  A resonable budget is about $150 per joint for the motors.  So there is $4000.     Then you have the structure,  Some of it would have to be metal but much could be 3D printed.  Then you would be about 50 quality ball bearing units, A computer,battery and whatever sensors are needed.   The total could be well under $16K

The problem with such a project is that a humanoid robot is a larger project than one person can do but the way most projects start is that one person makes a working product then other see it and join the project.    You need the working product before you have users willing to contribute back.

It wouold be great to find a way to break that chicken and egg problem.

Alan Timm

unread,
Feb 14, 2025, 1:50:17 PMFeb 14
to RSSC-List
Do you wanna see the G1 dance?

So this isn't the 16k version, but now that there are some number of G1s out in the field there's been alot of work on RL motion control policies for the G1 specifically.
This is clean demo.  The breakthrough shown here isn't mocap but translating that motion data into a motion plan or RL policy that can successfully execute on real world hardware.  Wow.


G1 dancing.jpg

Gmail

unread,
Feb 14, 2025, 1:58:53 PMFeb 14
to Alan Timm, RSSC-List
How do you know that there isn’t someone puppeting it?



Thomas

-  

Need something prototyped, built or coded? I’ve been building prototypes for companies for 15 years. I am now incorporating generative AI into products.

-

Need a great hardworking engineer? I am currently looking for a new job opportunity in robotics and/ or AI. 

Contact me directly or through LinkedIn:   


On Feb 14, 2025, at 10:50 AM, Alan Timm <gest...@gmail.com> wrote:


Do you wanna see the G1 dance?

So this isn't the 16k version, but now that there are some number of G1s out in the field there's been alot of work on RL motion control policies for the G1 specifically.
This is clean demo.  The breakthrough shown here isn't mocap but translating that motion data into a motion plan or RL policy that can successfully execute on real world hardware.  Wow.


<G1 dancing.jpg>







On Sunday, January 12, 2025 at 2:45:49 PM UTC-8 albertson.chris wrote:
Yes, $16K for a closed sourse remote control toy with no usable hands.   No hands is a big deal. Not programable is a big deal too.

What’s needed is a good open source humanoid project.   I think you could get the cost of entry way down if it were possible to build a partial robot, say a fixed stand with one arm and head-mounted sensors.  It could do many table top tasks.   Building out all four limbs and a torso would cost more, but could be done in stages.

A full humanoid has about 26 degrees of freedom if you do not count the hands.  A resonable budget is about $150 per joint for the motors.  So there is $4000.     Then you have the structure,  Some of it would have to be metal but much could be 3D printed.  Then you would be about 50 quality ball bearing units, A computer,battery and whatever sensors are needed.   The total could be well under $16K

The problem with such a project is that a humanoid robot is a larger project than one person can do but the way most projects start is that one person makes a working product then other see it and join the project.    You need the working product before you have users willing to contribute back.

It wouold be great to find a way to break that chicken and egg problem.



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

Chris Albertson

unread,
Feb 17, 2025, 10:34:41 PMFeb 17
to Gmail, gestalt73, RSSC-list


> On Feb 14, 2025, at 10:58 AM, Gmail <thomas...@gmail.com> wrote:
>
> How do you know that there isn’t someone puppeting it?

Try it. If you play back motion capturer to a robot, it falls over.

I don’t know what these guys did, but Boston Dynamics did share their method to get Atlas to dance. They captured human motion from a dancer and saved this as a “target motion”. Then Atlas did its best to try and reach those targets without falling over. To make it look good, the human dancer made moves he knew the robot could do. This is not literally “pupating” because the robot is not doing exactly what you did. It has a different center of gravity, arm and leg lengths, and so on. BD then use an MPC controller generate motion that best met all ther target points. They did say that the controller ran in a server room and connected to the robot over WiFi.

Tesla in their Robo taxi release had dancing robots too. They took a simpler route and used dance moves where the feet remained in a fixed position. Then, they bolted the feet to the floor so the robots were able to play back motion without falling over.

My guess is this G1 is using a technique with a level of sophistication between BD and Tesla. But still you can’t simply play back hard coded motion, the robot would fall over. Balance is always dynamic and can’t be precomputed.

(Well, it is dynamic unless you create a huge amount of static stability by bolting the feet to the floor or by making very huge feet, but is the robot still a “humanoid” if you do that? Bipedal walking with a dynamic balance is what has defined “human” for the last 6 million years, before we had speech and fire, going back to Lucy)

Alan Timm

unread,
May 1, 2025, 11:00:49 PMMay 1
to RSSC-List
Harrison from the SentDex youtube channel got himself access to a Unitree G1 Edu Ultra and is sharing his progress.
(for anyone else that wants to live vicariously through him.  :-) )

He did alot of neural network videos then took a break to touch grass.  Looks like he's back and he's hacking away at the G1.





Reply all
Reply to author
Forward
0 new messages