Unitree R1: From $5900

[Stephen Williams]

Getting there!  Whether this is interesting depends on how open the architecture is.  It could be a base for iterating on hands, head, sending, and ML control.  But if it is closed, with Unitree not supporting full access and iteration, without paying for the probably very expensive EDU edition, then it is mostly useless.

https://www.unitree.com/R1

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

[Thomas Messerschmidt]

So, it's an expensive toy? 

--
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/089055af-114c-4712-a384-3c891fc995ff%40lig.net.

[Peter N. Glaskowsky]

Not a toy, just a scam. Not real.

.         png

On Oct 6, 2025, at 21:10, Thomas Messerschmidt <thomas...@gmail.com> wrote:



To view this discussion visit https://groups.google.com/d/msgid/hbrobotics/CADyjTyaBt1PEWkeDG4wesjJuZ3tLxvWU-9YgWCRTySbmYuofXA%40mail.gmail.com.

[Stephen Williams]

I guess we will see if it is a scam or a toy.  Some of the specs seems like they wouldn't support what is in the videos, at all.  So there is that.

They have a number of warnings to not expect too much, but also still imply that it can balance, walk, etc.

sdw

[Chris Albertson]

As of today, I don’t see any documentation,  no user manual or API for the R1.   Unitree does have all of these for their other robots and you can download it from GitHub or Hugging Face..   Maybe the R1 is not quite ready to ship.

How open?  If it is like all their other robots, for $5,900 it will accept very low-level commands and a very, few high-level “demo” commands.

If Unitree does what they did in the past, the cheaper model has only “Sport mode” where it can do a short list of basic functions like Stand, Walk Forward, Stop, and a few others plus it will have some way to command each motor to a given angle and speed.  The edu model will have an Nvidia Orin computer installed and can be programmed at a higher level.  The cheaper model lacks the Nvidia controller computer.

The $5.900 model would be attractive to someone who can write their own software. (meaning that you can implement dynamically balanced movements) I’d say if you already have their G1 doing useful behavior in simulation, then this new R1 might be a cheaper way to test it in a physical robot.  But only after Unitree posts the documentation and API.

That said, I don’t think I have ever seen an example of a legged hoby-robot doing dynamically balanced walking or running.   This kind of result seems to require the expertise found in a university lab or larger robotics company.

[Peter N. Glaskowsky]

The Unitree R1 videos are CGI overlaid on real-world video. The easiest problem to see is that the robot does not affect the grass it’s walking on. Also, the shadows are completely wrong. (In the video with the jogger going by, compare the jogger’s fuzzy-edged shadow to the robot’s crisp stencil-like shadow.)

And as Chris said, no robot at anything like this price point has ever achieved this kind of dynamic motion. From the cost and agility perspective, it certainly helps that the thing is only 1.2 m tall and relatively (suspiciously) lightweight (25 kg vs. 1.5 m and 89 kg for the electric version of the Boston Dynamics Atlas), but Unitree is claiming improbable advances from one generation to the next, and the company had a history of releasing fake videos for its previous-generation robots as well.

I don’t doubt they’ll eventually ship an R1 that can walk and shadow-box and so on, like their G1 does. But the company is fundamentally dishonest.

.              png

--
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/c2427e89-4a20-4ce9-b237-666223ce9672%40lig.net.

[Alan Timm]

Hey Peter, I was struggling with the release of the G1 at 16k similarly to how you are now with the R1 now at 6k.

I can tell you without reservation that the promotional Unitree R1 demo video is 100% real.  The behaviors are ones that have already been demonstrated on the R1.

https://www.youtube.com/watch?v=v1Q4Su54iho

How are they able to offer it at 6k?  beats me.  efficiencies of scale and mass production?  subsidized by the CCP?  who knows.

But the demo is real, the R1 structure and setup is almost identical to the G1, and like many of us I'm waiting for pricing on the version with hands and a jetson.

I keep throwing my money at the screen but nothing happens.  If it's anything like the G1 release, we'll start to see units shipped at the very end of the year.

And getting low-cost humanoids out into the real world is what we need to close the gap on physical intelligence.  this is the way.  :-)

Alan

[Steve " 'dillo" Okay]

On Tuesday, October 7, 2025 at 11:48:59 AM UTC-7 Peter N. Glaskowsky wrote:

The Unitree R1 videos are CGI overlaid on real-world video. The easiest problem to see is that the robot does not affect the grass it’s walking on. Also, the shadows are completely wrong. (In the video with the jogger going by, compare the jogger’s fuzzy-edged shadow to the robot’s crisp stencil-like shadow.)

There's definitely some "tells" going on in that video WRT depth-of-field, shading, etc.
I missed the grass but kept noticing the flat Gouraud-shaded surfaces on the body. 

 

And as Chris said, no robot at anything like this price point has ever achieved this kind of dynamic motion. From the cost and agility perspective, it certainly helps that the thing is only 1.2 m tall and relatively (suspiciously) lightweight (25 kg vs. 1.5 m and 89 kg for the electric version of the Boston Dynamics Atlas), but Unitree is claiming improbable advances from one generation to the next, and the company had a history of releasing fake videos for its previous-generation robots as well.

Since moves and agility in a robot are really just a visual proxy for motors being able to move to a certain position in a certain amount of time, I believe that the moves are real. 

They might be scripted, but given the right number of motors with the right specs, it's possible to script a robot to do that. 

I can do some amazing moves on my skateboard and snowboard because I only need to have my body achieve specific positions for a split second before moving on to the next. 
I want to see one of these robots stand "en pointe" or hold a "Tree" or "Warrior III" pose for 30 seconds, which would strain the motors and force it to do some serious PID control

and/or computation. 

I don't need another robot to demo or impress people with. I've got plenty of those already. 

Owen Lars, Luke Skywalker's uncle on Tatooine said it best:

"I need those droids up on the South Ridge working on those condensers by mid-day or there'll be hell to pay"

'dillo

[Chris Albertson]

>
> Since moves and agility in a robot are really just a visual proxy for motors being able to move to a certain position in a certain amount of time, I believe that the moves are real.
> They might be scripted, but given the right number of motors with the right specs, it's possible to script a robot to do that.

You cannot “script” a humanoid robot to stand upright on two feet while moving its arms. The reason is that the robot’s movements are never 100% perfect. So each time you run the motion script, the performance will be “off” by a very tiny amount, and that is enough to unbalance the robot and have it fall over. A simple script is what we call “open loop” control, and it absolutely would never work for humanoid walking on a natural surface like grass. You need closed-loop control.

In fact, closed-loop is not enough. You can depend on something very simple like a PID algorithm, to close the loop because balance many times to not be directed to a current set point. In other words, we have to allow imbalance if it leads to a further balance. Running is that, say, we tolerate a fall of the rear foot because we plan a future recovery. In robotics, we use simulation-based optimized to minimize some cost function at some near-future time. This is pretty much “just how it is done”. So you can never script this because you can’t know in advance the exact rotation and speeds or even the location of the surface or how soft each bit of grass will be. It is a reasonably hard problem, but it is also well studied.

In practical programmer terms — The walking and even standing routine needs inputs from the IMU and other sensors. We can’t know those values in advance.

In any case, as I wrote before, I have not seen even one case of a hobby/enthusiast programming a dynamically balance robot to walk, run, or do anything on a natural surface. Such feats of programming are done by larger companies or university labs.

>
> I can do some amazing moves on my skateboard and snowboard because

You can do this because your brain has dynamic balance and motion planning hard-wired in without you having to think about it. If we were to record your motions andplay them back on a robot, it would fall over. Boston Dynamics has done this many times and shown us how it works. They motion capture a dancer and then play it back at TARGET points where the targets are just that, suggestions to the robot about where to aim. The robot then does its best to reach those targets, but reaching too is secondary to not falling down. So the end result is a not-exact playback of the motion capture. An exact playback would be easy, but the robot would quickly fall down.

[Stephen Williams]

That said, I don’t think I have ever seen an example of a legged hoby-robot doing dynamically balanced walking or running.   This kind of result seems to require the expertise found in a university lab or larger robotics company.

These look promising as starting points:

https://www.instructables.com/Arduino-Controlled-Robotic-Biped/

https://arxiv.org/html/2502.00893v1

https://www.ce.cit.tum.de/ics/news/article/exciting-news-openwalker-humanoid-robot-walking-software-now-available/

https://arxiv.org/html/2406.01152v2#S7

Stride: https://arxiv.org/html/2407.02648v1

https://github.com/loco-3d/crocoddyl

Related to crocoddyl:

https://stack-of-tasks.github.io/

https://humanoid-path-planner.github.io/

https://github.com/stack-of-tasks/pinocchio

scienteer

--
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/78017E8F-08B6-434C-A2BC-626DECCC86AE%40gmail.com.

[Peter N. Glaskowsky]

Alan,

I am equally certain that there is not one scene in the video you link to that depicts real-world movement of the robot.

I can point to obvious evidence of fakery, like the fact that the scene at 0:00:18 mixes CGI grass with real grass to help reinforce the illusion of reality but doesn’t even do collision detection in the CGI domain, so when the robot pops out of the handstand, its right hand moves right through some CGI grass without disturbing it at all.

But if that doesn’t convince you, there’s probably nothing I can say that will, and I won't spend any more time trying.

Best,

.              png

To view this discussion visit https://groups.google.com/d/msgid/hbrobotics/f65e797a-d530-44f6-943c-8728ff4f24b7n%40googlegroups.com.

[Dan]

I guess you missed the demo of WATSON my full sized biped walking on a table at the HBRC meeting about 8 years ago .

https://groups.google.com/d/msgid/hbrobotics/f704733e-8a59-48ce-a642-6a461aa595c0%40lig.net

.

[Stephen Williams]

Hey man, what have you done for me lately? ;-)  Do we have a video of that?

We need to get to the nimble monkey walking, dancing, jumping stage.  Seems possible with current technology + a lot of improved code.

To view this discussion visit https://groups.google.com/d/msgid/hbrobotics/1021546559.951252.1759954664129%40mail.yahoo.com.

[Dan]

Chris Mayer and I have started a new project.  Unlike Unitree or even the K1Kscale open project, this one is a full sized biped design that is completely 100% open and published design. Anyone that is interested in contributing should contact us privately.

To view this discussion visit https://groups.google.com/d/msgid/hbrobotics/58a09e88-2b2a-4d79-af82-1da3ebd429c3%40lig.net.

[Stephen Williams]

I am interested.

I have several parts that I'm working on.  I haven't decided whether I will make all of it 100% open/free in every sense, but everything will certainly be open + free (or nearly free) for any hobbyist.  I may just make everything completely free.

I'm working on a new hand now.  That has a higher likelihood of being free or mostly free.  Perhaps I might just charge a little for the printable model?  I'm not likely to actually innovate much on that, just remix ideas so that something is really easy & cheap to build.

I now have the computer to do simulations + RL for motion control, planning.  Especially when I have interesting hardware finished enough.

I have some good maker capabilities now, including cutting & welding metal.

I am still short on time, but getting finished with a big project.  And that project is relevant too.

sdw

To view this discussion visit https://groups.google.com/d/msgid/hbrobotics/185322499.938758.1759956060436%40mail.yahoo.com.

[Thomas Messerschmidt]

I have been trying to build a hand for my fembot. Hard to build a slender hand. 

Thomas

On Oct 8, 2025, at 1:53 PM, 'Stephen Williams' via HomeBrew Robotics Club <hbrob...@googlegroups.com> wrote:



To view this discussion visit https://groups.google.com/d/msgid/hbrobotics/6c73354e-7a17-42f6-89df-753a6b3c86f3%40lig.net.

[camp .]

    Paraphrasing Andra Keay here: Humanoid robots are never going to get any worse than the current state of technology; they are going to continue to improve. They're going to become more capable and become more affordable. They're here to stay because the range of use cases for them is pretty broad.

Enjoy,

Camp

To view this discussion visit https://groups.google.com/d/msgid/hbrobotics/185322499.938758.1759956060436%40mail.yahoo.com.

[Chris Albertson]

Open Walker comes close, but they never go out of a stable balance.

Walking robots have the concept of “ZMP or Zero Moment Point.”   It is like Center of Gravity (CG) but for moving objects where the force is more than just gravity and accounts for things like momentum and acceleration.    Like CG, if you keep the ZMP inside of the  supporting base, the robot does not fall over.   This covers movements like stopping fast and driving around corners.   Just keep this point inside the supports, and you stay upright.

A simple walking robot will have a tight loop servo, maybe a PID that will always do what it takes to keep the ZMP inside the boundaries of its feet.   The two-wheeled Segway-like balance bots do the same.    If you can write the code to never allow the ZMP outside of the feet, the ‘bot will never fall.   This allows the humanoid robot to lift a tray by leaning back to keep the center of gravity over the feet.   It is actually not hard to write this kind of code.    It is just a tight loop that runs at about 50Hz and moves the waist, ankles, or knees to add or remove a bias to maintain balance.

But humans do not walk like that.  We intentionally place our ZMP outside of our feet and start to fall over, but we have a plan to place the other foot in such a way to break the fall, then we do it again.    Our walking algorithm is fundamentally different in that, unlike a simple PID, we plan our future and will allow a very significant unbalance if the future reward is great enough.  We optimize cost at some future time horizon.  A PID, on the other hand, only removes past error, not future error.    

So, you cannot get to a point where you are optimizing the future but making code improvements to a PID-like servo.   You have to toss all your code in the trash and start over.     If the robot wants to deliver that beer to his master quickly, the robot needs the rear foot to be as far to the rear as it will go and the front foot as high in the air and as far forward as it can go.  Obviously, the robot is maximizing its imbalance so as to use gravity to generate a torque.   You will NEVER get there if you start with a servo’d balance.

People and animals try to minimize future error while “balance serves” look only at the recent past errors.   They are so very different that you cannot modify one to get to the other.

No one does this really well yet.   The bigger players are impressive but human-like goal-directed behaviours are still not there.

If the video of R1 is real then they would be using the same algorithm as Boston Dynamics did for Atlas.  They motion captured a human dancer and used these points as targets for an MPC-based controller.

With that background, what I say no single enthusiast has mastered dynamic balance, I mean mastered a future-looking algorithm.     All the PID-like ones result in a walk that looks like a shuffle, K=look at the Open Walk video.  The robot contorts its body to keep ZMP hover its feet.   Then look at a Boston Dynamics video where Altas is in constant free fall but prevents falling with well-planned footwork that is computed in real-time.

To view this discussion visit https://groups.google.com/d/msgid/hbrobotics/f704733e-8a59-48ce-a642-6a461aa595c0%40lig.net.

[Thomas Messerschmidt]

And it's going to take 10-20 years to create a "Rosie" or a "C3P0" that can be as physically capable as a human. It's a VERY large technological jump from a robot that can do a single task well to a robot that can do any task well, especially bipedal humanoids. 

To view this discussion visit https://groups.google.com/d/msgid/hbrobotics/1491009320.975448.1759960887048%40mail.yahoo.com.

[camp .]

> And it's going to take 10-20 years to create a "Rosie" or a "C3P0" that can be as physically capable as a human.

Twenty years pass quickly.

- Camp

To view this discussion visit https://groups.google.com/d/msgid/hbrobotics/CADyjTyak9%3D%2BV8vGxcGwGc2tXe-LGXAnmjcumij_j_p8au2Un_w%40mail.gmail.com.

[Thomas Messerschmidt]

True. 

And did I mention safety? Any one of those humanoid robots can easily hurt you, even cut off a finger. If a robot is as strong as a human, then for sure, accidentally or otherwise, it can kill a human. Safety may be the trickiest part of wide spread humanoid robots, especially in the home. 

Thomas

To view this discussion visit https://groups.google.com/d/msgid/hbrobotics/1719816035.1018755.1759969418490%40mail.yahoo.com.

[Chris Albertson]

>
> And did I mention safety? Any one of those humanoid robots can easily hurt you, even cut off a finger. If a robot is as strong as a human, then for sure, accidentally or otherwise, it can kill a human. Safety may be the trickiest part of wide spread humanoid robots, especially in the home.

You can say the same about self-driving cars. Like a Waymo taxi. A car has the potential to injure not only the people inside but also pedestrians and other road users. But after millions of miles of data, we see the probability of a Waymo taxi causing harm is 4 or 10 times less than for a human driver.

Yes, I know, cars do not drive in your kitchen. Domestic robots do different tests than robot taxis. But I think of rules that might help, like not moving at all if a human is nearby. Just stand motionless. Yes, but they might fall over and crush a child or a pet. There might be even safer behaviors than standing.

Safety will be an issue. You are right. But I think a good failsafe multi-layer safety system might be possible. The car example is real. Robots can be safer.

[Thomas Messerschmidt]

Currently it’s “The Wild West” when it comes to safety and humanoid robots.


Thomas



On Oct 8, 2025, at 7:03 PM, Chris Albertson <alberts...@gmail.com> wrote:



--
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/92D9B679-56B1-4E83-9B53-43592943A638%40gmail.com.

[Stephen Williams]

You misspelled that: It is the Wild WestWorld.

sdw

[Thomas Messerschmidt]

😆

Thomas

On Oct 8, 2025, at 10:31 PM, 'Stephen Williams' via HomeBrew Robotics Club <hbrob...@googlegroups.com> wrote:



To view this discussion visit https://groups.google.com/d/msgid/hbrobotics/d6dd9e30-9317-46d3-a55f-2906da97196c%40lig.net.

[Tom Bertalan]

> at 0:00:18 mixes CGI grass with real grass to help reinforce the illusion of reality but doesn’t even do collision detection in the CGI domain, so when the robot pops out of the handstand, its right hand moves right through some CGI grass without disturbing it at all. 

Peter, Look closer, in slow motion. grass is clearly moved with every sweeping foot or hand movement, including the big tuft in front.

> In the video with the jogger going by, compare the jogger’s fuzzy-edged shadow to the robot’s crisp stencil-like shadow.

If this is the same scene on grass, note that the jogger is on a slope slightly away from the camera, so his shadow is seen obliquely through grass, which creates a fuzzing effect. The robot's shadow is nearer to us and on ground more directly tilted towards the camera.

>  (suspiciously) lightweight (25 kg vs. 1.5 m and 89 kg for the electric version of the Boston Dynamics Atlas)

I'd put it down to crappier materials and tiny battery. BD has B2B relationships they're taking seriously, but this R1 is a technology demonstrator not intended for actual use.

> No one does this really well yet.   The bigger players are impressive but human-like goal-directed behaviours are still not there.

Chris, BD pioneered whole-body trajectory optimization (WBTO), but other players have since leapfrogged them on implementation (Agility comes to mind). SOTA is large vision/action models, where the action is at the abstract level of "go here, grab that" and so provides a higher-level target for lower-level WBTO (which itself cascades to simpler high-frequency motor torque control loops).

Also, just because you're doing WBTO doesn't mean that ZMP balancing is useless. (1) keeping CoM over footprint can be another objective in the TO, and (2) if you're just standing still, you can save computation power by switching to PID balancing.

Alan, 1000% subsidized by government, yeah.

[Dave Everett]

I think this demo suggests one question.

How often do you need a somersaulting robot?

Dave

[Dan]

How many licks can it take before it breaks...1....2....3.

Three, then you need replacement parts.

Oh yeah....Unitree parts are not available for purchase.

How do I know? I have had some on order for 4 months.

Good Luck!

--

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/CAG61pbcG%3DPzFrVssLkU1QzT%2B_GPorWdbRcwG6DKfZDY8GuUO-A%40mail.gmail.com

.