The economics of the cheapest humanoid currently available (Unitree R1)

[Alan Timm]

Thomas started a conversation with me today about the Unitree R1 and it was so good I'm going to share the highlights.  (Thanks Thomas!)

https://www.unitree.com/R1

https://shop.unitree.com/products/unitree-r1?variant=50941270163689''

So Unitree updated their product and storefront page with pricing and details on the 3 different R1 models.

A little back of the napkin calculations:

The R1 Air and R1 are radio controlled toys.  you can drive them around but you can't program them.

And all of their robots have enabled secureboot, so it's unlikely you'd be able to buy the cheap one and crack it to unlock programming capability, but...

The same guys that did the jailbreak for the Go2 are on the case, so you never know?

And for those of you thinking you can buy the cheap one and use it as a bag-o-parts and make your own control systems and cpu --  first you'll need to crack the encrypted comms for the bldc motor units.  to my knowledge that hasn't been done on the go2 motors either.

The R1 EDU is the only version that is programmable, but we're talking about company/university/committed hobbyist levels of programming.  Not blockly so much as end-to-end programming like gr00t or isaac lab.  not for the faint-of-heart.

So assuming that EDU is the 100TOPS Jetson Orin NX 16gb unit, you're not going to be able to run Gr00t on it anyways so you're limited to other frameworks and issac lab rl trained models.   Just like the G1, people are throwing compute backpacks on these guys with either Orin AGX or Thor, or just doing the compute remotely.

And instead of paying for their incredibly overpriced DEX-3 hands you can go with your own open source hand design or something like the amazinghand:

https://cad.onshape.com/documents/95115c05528ff95759d5a625/w/a17974ee809ad259cf346b79/e/b78d99e19a5237928e2e9c57?renderMode=0&uiState=6a383790654e2233d12ec8c3

So even at $16k it's the lowest cost full-size-ish humanoid on the market right now.

But you're going to be spending most of your time with the R1 under a gantry anyways until you figure everything out.

So what do you think?  

[Thomas Messerschmidt]

So I’m talking to Alan‘s AI?

Thomas Messerschmidt

-  

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

Contact me directly or through LinkedIn:   

https://www.linkedin.com/in/ai-robotics/

On Jun 21, 2026, at 12:25 PM, Alan Timm <gest...@gmail.com> wrote:

Thomas started a conversation with me today about the Unitree R1 and it was so good I'm going to share the highlights.  (Thanks Thomas!)

https://www.unitree.com/R1

https://shop.unitree.com/products/unitree-r1?variant=50941270163689''

So Unitree updated their product and storefront page with pricing and details on the 3 different R1 models.

A little back of the napkin calculations:

<screenshot_20260621_120508.jpg>

The R1 Air and R1 are radio controlled toys.  you can drive them around but you can't program them.

And all of their robots have enabled secureboot, so it's unlikely you'd be able to buy the cheap one and crack it to unlock programming capability, but...

The same guys that did the jailbreak for the Go2 are on the case, so you never know?

And for those of you thinking you can buy the cheap one and use it as a bag-o-parts and make your own control systems and cpu --  first you'll need to crack the encrypted comms for the bldc motor units.  to my knowledge that hasn't been done on the go2 motors either.

The R1 EDU is the only version that is programmable, but we're talking about company/university/committed hobbyist levels of programming.  Not blockly so much as end-to-end programming like gr00t or isaac lab.  not for the faint-of-heart.

So assuming that EDU is the 100TOPS Jetson Orin NX 16gb unit, you're not going to be able to run Gr00t on it anyways so you're limited to other frameworks and issac lab rl trained models.   Just like the G1, people are throwing compute backpacks on these guys with either Orin AGX or Thor, or just doing the compute remotely.

And instead of paying for their incredibly overpriced DEX-3 hands you can go with your own open source hand design or something like the amazinghand:

https://cad.onshape.com/documents/95115c05528ff95759d5a625/w/a17974ee809ad259cf346b79/e/b78d99e19a5237928e2e9c57?renderMode=0&uiState=6a383790654e2233d12ec8c3

So even at $16k it's the lowest cost full-size-ish humanoid on the market right now.

But you're going to be spending most of your time with the R1 under a gantry anyways until you figure everything out.

<screenshot_20260621_122126.jpg>

So what do you think?  

--
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 visit https://groups.google.com/d/msgid/rssc-list/50eb4357-b0f2-4bee-aa7e-73bf713ec77dn%40googlegroups.com.

<screenshot_20260621_122126.jpg>

<screenshot_20260621_120508.jpg>

[Alan Timm]

Not yet.  Lloyd has access to my email and calendar but isn't allowed to respond or write them -- yet.  :-)

Alan

[Alan Timm]

Oddly enough there isn't alot on youtube showing off the R1 yet.

Here are the only two substantial videos I can find.  Both are non-english but give you a good look at the robot and capabilities.

https://www.youtube.com/watch?v=wy0Kv3TYdnM (Czech Republich)

https://www.youtube.com/watch?v=-E29jx43zcA (Russia)

[Thomas Messerschmidt]

It's interesting to watch the videos. I agree, these are expensive toys. 

A couple of observations:  I don't like the robot's head. It looks like a skull with ski goggles. What were they going for? And, it is much smaller than I imagined. 

And the quadruped--the guy sat on it! I didn't know those could take a man's weight. 

To view this discussion visit https://groups.google.com/d/msgid/rssc-list/64eff351-1d96-4edd-a3ed-23102d70cdb0n%40googlegroups.com.

[Alan Timm]

Yeah they really missed the mark with the R1 head and the H2 head in my opinion.  

So weird.

So I probably won't be plopping my credit card down for any of these any time soon.

And I was thinking...  

Alfie is almost as tall as R1, has working hands, and has a propensity to fall over just like the R1.  And at 1/10th the price and supports the full gr00t n1.7 model on board.

I think Alfie will keep my busy for a while.  :-)

Alan

[Alan Timm]

It looks like there were a couple of revisions of the head, and the settled on the worst one.

Here is most likely an earlier version.

https://www.robotshop.com/products/unitree-r1-edu-u2-humanoid-robot

[Thomas Messerschmidt]

What will the groot model do for the robot? Have you gotten any of it working? I have yet to dive into the capabilities of it.

Thomas Messerschmidt

-  

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

Contact me directly or through LinkedIn:   

https://www.linkedin.com/in/ai-robotics/

On Jun 23, 2026, at 10:31 AM, Alan Timm <gest...@gmail.com> wrote:

Yeah they really missed the mark with the R1 head and the H2 head in my opinion.  

To view this discussion visit https://groups.google.com/d/msgid/rssc-list/e38c973e-e0fd-47e5-9d37-b42ebc129a55n%40googlegroups.com.

[Chris Albertson]

>
>
> The R1 Air and R1 are radio controlled toys. you can drive them around but you can't program them.

I think you might consider them as a box of robot parts. Buy the robot, then go in with a screwdriver and remove the computers, but keep the motors and the H-bridge drivers and the battery. I think it would actually be cheaper to do this hardware hack than to buy all the motors and machine the metal parts.

Don’t bother to try and hack the software; replace the computers.

That said, 99% of the people who would buy this would be happy to simply drive it around using an Xbox controller.


But if you really want to do some programming, you can’t start with big projects like generalized housekeeping, construction tasks or building a spaceport on Mars. You have to start with simple things like “grasp planning,” where the robot decides how it will place its fingers over some object. For that, all you need is a robot hand on the end of a stick. There is so much of this simple stuff, and in each case, all you need is enough hardware so you can validate the algorithm. It would be years before you could put it all together.

A fully programmable robot will do you no good if you don’t have the software you need to program it.


Every time I price out the components needed to build a full-size humanoid that is powerful enough to do real tasks, it ends up costing more than the price of an entry-level Toyota. You can’t do it for only $20K. So seriously, buying a cheap robot, then adding your own computer, is actually cost-effective.

[Alan Timm]

Hey Chris,

You wouldn't be the first person to think that, and there's a handful of very motivated and talented peeps attempting to figure out how to crack the R1 like we did with the Go2, but there's a few things getting in the way.

Looking back to the Go2, more than one person bought the Go2 air specifically to re-purpose the motors for other projects, but were unable to interface to them as they had no documentation on the proprietary comms for the motors.  Both the protocol and the CRC got in the way.  Even after we Ghidra'd some of the earlier firmware to figure it out.  To my knowledge no one was able to get the motors to work outside of the robot.

Then later on they pushed updated firmware to all Go2 motors as part of a firmware update and the comms changed again.   Then they enabled secure boot and started encrypting and obfuscating the host firmware to make it almost impossible to figure out. 

AS OF RIGHT NOW, it appears that R1 security is just like early Go2 in that some portions aren't as locked down as they'll end up in the near future.

I'd keep your credit card in your wallet for a bit, if it's anything like Go2 the most you can hope for is to gain root to the host computer and go from there.

And if they do figure something out and you're feeling lucky?

You could buy the cheap versions, R1Air @ $8k or the R1 Standard @ $10k and hope yours was manufactured and shipped before they locked everything down.

I'll definitely post updates as they (hopefully) make progress.

Alan

[Chris Albertson]

The security features are totally unimportant. And not even relevant here.

If you remove enough of the electronics eventually to have three wires leading to each motor and they are analog. The PCB that controls the motor would be tossed in the trash.

The goal should be to not leave even one byte of firmware in the robot.

If you are even taking about “gaining root access to the main computer,” you missed my point. Toss the main computer in the trash and don’t bother. Computers are cheap as dirt. Buy a new one.

I would never in a million years want a robot if I did not also have 100% of the source code that runs it. The SimpleFOC project provides very good open-source motor control and many reference hardware designs.

I know 100% that one ESP32-S3 microcontroller can run FOC computation for two motors and also run a CAN bus interface. To do that, you need a generic gate driver chip and 6 MOSFETs, and you need to hire a company like JLCPCB to make it for you. They are cheap and charge about 20 cents for the PCB, but those high-power MOSFETs are not cheap, costing maybe $20 for all the parts.

These computers are cheap to replace. Seeed Studio sells them for $5 each. You don’t hack it, you replace it. Hacking might be impossible because there are one-time on-chip fuses that can be blown. But it does not matter because your interface is to power leads on the motor.

[Chris Albertson]

To be clearer. See the photo below of an example motor. The software hacker looks at that USB port and tries to figure out what commands he can send to that port. He fails because the commands are encrypted.

The hardware hacker sees four screws securing a cover plate and removes the plate, then removes the PCB. Then he puts his own PCB in its place and puts the cove plate back on. To him, the encryption is irrelevant. This works because the PCB costs only about $12 to replace.

[Alan Timm]

Quick update.

A buddy of mine just took delivery of his Unitree R1 Pro and showed me his import tariff bill.

My best guess for import tariffs was $2360 and that's less than 1/2.

I think that puts his total out of pocket for the Unitree R1 Pro at ~$8170.00?

[Chris Albertson]

And for those of you thinking you can buy the cheap one and use it as a bag-o-parts and make your own control systems and cpu --  first you'll need to crack the encrypted comms for the bldc motor units.  to my knowledge that hasn't been done on the go2 motors either.

If the Unitree engineers were even halfway competent, “cracking” would be literally impossible by anyone.   Modern encryption is very powerful.   The only solution is to trash the controller PCBs and replace them with Open Source designs.  Fortunately, there are good open-source motor controllers

But you're going to be spending most of your time with the R1 under a gantry anyways until you figure everything out.

So what do you think?  

What do I think?     This would be a very stupid development plan. Why would anyone buy a $8000 machine and then spend a year or two trying to write software to make it work?

Wouldn’t it be better to FIRST spend a year or two writing software and then later buy some hardware?  Not only does this kick that high cost down the road, but the hardware will likely be both better and cheaper if you wait.

[Alan Timm]

Hehehe, that would be the smart thing to do.  And sooooo much cheaper.

If you have an nvidia card?  Isaac-Sim and Isaac-Lab.

If you don't?  Mujoco is amazing.

For anyone interesting in exploring training with RL or gr00t policies under simulation, here's the urdf file for the R1

https://drive.google.com/file/d/1g8DTN_KrHNHjrEIEI8_wnxLVy8-1KBPf/view?usp=sharing

As of right now the R1 hardware platform is mostly not locked down, but that could change at any time.

There's a bunch of smart peeps on Discord right now figuring things out.  Here's a link if you're interested.  https://discord.gg/PePD9Ra8

Alan

[Chris Albertson]

On Jul 20, 2026, at 5:35 PM, Alan Timm <gest...@gmail.com> wrote:

Hehehe, that would be the smart thing to do.  And sooooo much cheaper.

If you have an nvidia card?  Isaac-Sim and Isaac-Lab.

If you don't?  Mujoco is amazing.

Mujoco is open source.  Isaac is only “free to use”.      Well, nearly free becuase to need one of these to use it

As long as you are simulating, why not sim the G1?    The G1 is a much more capable robot than is the R1.  It has features that were designed to make it usable for doing things with its hands and arms.

R1 has 5 degrees of freedom arms while G1 has over-articulated 7 degrees of freedom in each arm.    This means R1 would have a hell of a time using a screwdriver.

The problem is that a 5-DOF arm cannot simultaneously control position in (x,y,z) and orientation in (roll, pitch, yaw).  If you count “x, y, z, roll, pitch, yaw,” you get 6.    The R1 arm can control any five of those, but the other will need to be some random result that needs to be corrected by odd torso leans and twists.  Can it do that and still balance over its feet?   I don’t know, you have to try in simulation.

Both humans and the G1 have 7-DOF arms, so we actually have an infinite number of solutions for every position and angle of the hand.  Having an infinite number of solutions allows us to do things like avoid making contact with the wall while turning the screwdriver; we naturally know to move the elbow to avoid it.  With so many solutions, one is likely to be better.

R1 and G1 are about the same (except for size and power) if all you want to do is walk around.   Both have 6-DOF legs.  The difference is in the arms and waist.

So, I’ve pretty much decided I’ll simulate a G1.  But files are free to download.    Then later, if I want to move to hardware, I can edit the file to match whatever the hardware design happens to be.     Files are easy to edit so I can try stuff out and see if the idea works.

Oh, about the above, you could always rent a GPU.  It is free for the first so many hours.  After a few months, you’d know how much compute power you really need.

--
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 visit https://groups.google.com/d/msgid/rssc-list/58f7a525-88c2-4b58-b456-30e95b3c6289n%40googlegroups.com.

[Thomas Messerschmidt]

Isn’t physical training less cpu/gpu intensive than text, video, audio, etc. training? I mean, there’s much, much less data, right?

Thomas Messerschmidt

-  

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

Contact me directly or through LinkedIn:   

https://www.linkedin.com/in/ai-robotics/

On Jul 20, 2026, at 5:35 PM, Alan Timm <gest...@gmail.com> wrote:

Hehehe, that would be the smart thing to do.  And sooooo much cheaper.

--

[Chris Albertson]

> On Jul 21, 2026, at 5:25 PM, Thomas Messerschmidt <thomas...@gmail.com> wrote:
>
> Isn’t physical training less cpu/gpu intensive than text, video, audio, etc. training? I mean, there’s much, much less data, right?

Yes. The big difference is the size of the model. An LLM, even a “tiny” one, might have billions of parameters, whereas a robot location model might have maybe a million at best or much smaller. The LLM could easily be 100,000 times larger

Training an LLM is a data center-sized project. Training a robot only needs one big computer.


BTW, Thomas, have you decided on a plan for humanoid robots?


> Thomas Messerschmidt

[Alan Timm]

Based on feedback from a few folks in the USA that have ordered R1s, 

here's an updated cheat sheet of final costs after reducing the import tariffs based on what they were billed.

The R1 Air doesn't have the 2 axis neck nor the 2 axis waist, and the EDU unlocks development and includes an orin of some sort.

They've been ordering the R1 standard.

[Thomas Messerschmidt]

So are you planning on buying one. I know you are tempted!

Also, Ive joined a small team of DIY folks and I am working towards developing my own. 

Thomas Messerschmidt

-  

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

Contact me directly or through LinkedIn:   

https://www.linkedin.com/in/ai-robotics/

On Jul 29, 2026, at 8:32 AM, Alan Timm <gest...@gmail.com> wrote:

Based on feedback from a few folks in the USA that have ordered R1s, 

here's an updated cheat sheet of final costs after reducing the import tariffs based on what they were billed.

<screenshot_20260729_082729.jpg>

The R1 Air doesn't have the 2 axis neck nor the 2 axis waist, and the EDU unlocks development and includes an orin of some sort.

They've been ordering the R1 standard.

On Tuesday, July 21, 2026 at 10:25:25 PM UTC-7 albertson.chris wrote:

> On Jul 21, 2026, at 5:25 PM, Thomas Messerschmidt <thomas...@gmail.com> wrote:
>
> Isn’t physical training less cpu/gpu intensive than text, video, audio, etc. training? I mean, there’s much, much less data, right?

Yes. The big difference is the size of the model. An LLM, even a “tiny” one, might have billions of parameters, whereas a robot location model might have maybe a million at best or much smaller. The LLM could easily be 100,000 times larger

Training an LLM is a data center-sized project. Training a robot only needs one big computer.


BTW, Thomas, have you decided on a plan for humanoid robots?


> Thomas Messerschmidt

--
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 visit https://groups.google.com/d/msgid/rssc-list/3b84c740-65c2-48db-a5d4-1d65e80b4c17n%40googlegroups.com.

<screenshot_20260729_082729.jpg>

[Chris Albertson]

> On Jul 29, 2026, at 9:15 AM, Thomas Messerschmidt <thomas...@gmail.com> wrote:
>
> So are you planning on buying one. I know you are tempted!
>
> Also, Ive joined a small team of DIY folks and I am working towards developing my own.

What is the plan for this? Is it a full-sized robot like Optimus or Atlas or something scaled down. Do you all have a roadmap?

A lot of us would like to at least read along and watch this process.


To anyone who wants to buy one, I have the same question: What is the process after buying, what is the goal. Again, I think you will find a following if you posted frequent updates.


In both case, set up a Youtube channel and if it is good enough it could help pay for the robot.

[Alan Timm]

As tempting as it is...

Alfie can do everything the R1 can as long the the ground is level, including a propensity to tip over.

We're all working with the same software toolkit for the most part.

:-)