That's a cool little robot, and I see the appeal, definitely as an educational toy. I don't think the fidelity is there for real autonomous research. But the bigger issue imo is that there's no way this should cost $2000.
You've got a $250 computer, some lidar+camera sensor for maybe $1-200, 6 servos, and cheap plastic. Plus you want to charge a $50/mo software subscription fee for some software product, whatever I guess that's beside the point.
No shade on the idea because low-cost robotics is an unsolved need for the future. But this current iteration is just not competing well with other alternatives. Perhaps this is more of a comment on what we can accomplish in the West vs what's possible in Asia.
Actually the BOM cost required to make something stable that can execute manipulation tasks well enough is around $1k+ hence our price. You will find very cheap robots that can pretend to do what this one can, but in practice won't work well enough.
As for the unitree robot, this one is not unlocked for development, does not have onboard GPU, and does not have an arm. If you want it, check the price they give, it's very prohibitive.
You could attach a cheap arm to it but it would also not be stable enough for AI algorithms to run it. We're researchers ourselves, we would have made it cheaper if we could, but then you just can't do anything with it.
Our platform will deliver the experience of a real AI robot, anything cheaper than that is kind of a lie - or forces you to assemble and calibrate, which we do for you here. It is just the nature of trying to deliver a really complete product that works, and we want to stand for that.
As a side note, the previous generation of research platforms for that size made in Asia were the Turtlebots, which go for that same price, but without GPU, arm...
I would say the problem is that most manufacturers, including chinese, sell you platforms that are not reliable enough for AI manipulation, and there's a race to the bottom for it, to which we try not to participate to
Does the MARS hardware really remove the hidden extras (computer with a gpu) mentioned as the downside of HF SO-101 or LeKiwi? While a jetson is good for inference, I feel like to train VLAs you would need access to a powerful GPU regardless. For Lerobot based hardware training ACT was relatively low profile if you use low resolution for the camera feeds, but with increased resolution or with more than one camera I already saw needing more than 8GB of VRAM. If VLA is on the table, finetuning something like the open sourced version of pi0 should already necessitate access to more than one 4090 or above I think.
Also, do you have plans for community-level datasets? I think Lerobot sort of does this with their data recording pipeline and HF integration.
One of our objectives here was to fix everything that we don't like about the SO-101 and the Kiwi, which have several hardware and software flaws in our view. Including, yes, the constant need for a computer to simply run your robot.
The training does require external GPUs (but we provide that infra for free, straight from the app!), but the onboard jetson can run models trained though, as you can see in the examples. Everything you see in the vids is running onboard when it comes to manipulation, because we use a special version of ACT made specifically by us for this robot, that also includes a reward model (like DYNA does).
We have developed this system to also be able to run the other components smoothly so it also does SLAM, and has room for more processing even when running our ACT.
Now indeed this cannot run Pi-0 but from our experience - and the whole community in general - VLAs are not particularly better than ACT in the low data regime, and need a lot more compute.
As for community-level datasets, yes this is the plan. Anything you train can already be shared with others - just share the files. We didn't develop a centralized place for sharing datasets and behaviors but it is on the plan.
> BASIC is accessible for free to all users of Innate robots for 300 cumulative hours - and probably more if you ask us.
Is BASIC used just to create the behaviours or to run them too? It sounds like this is an API you host that turns a behaviour like "pick up socks" into ROS2 motor commands for the robot. Are you open sourcing this too, so anyone can run the (presumably GPU heavy) backend?
Does the robot needs an internet connection to work?
Also, more importantly, what does it look like with googly eyes stuck on?
On BASIC: Yes it does require an internet connection and until we figure out how this works for you it will remain free of use!
It is required to run them, not to create them. And it's not about running "pick_up_socks", this one can already run on your robot. BASIC is required to chain it with other tasks such as navigating to another part of your house and then running another skill to drop the sock somewhere for example
Thank you for the remark, we will make it clearer in the docs
As a consequence: The robot does not necessarily require Internet to run, but if you want it to chain tasks while talking and using memory, yes it does.
and much higher interfaces for interaction and ai manipulation. Like directly recording episodes of training data so that the arm can use a VLA instead of simple IK.
this website froze my computer for several minutes, like i'm browsing the web in 2004. making the impression that the robot will work the same way, very uninterested now
ain't got time for that. happened in both firefox (locked down with lot of privacy settings/adblockers) and chrome (no privacy/adblockers etc). froze entire computer on both browsers for a few minutes for each browser. my computer has like 16gb of RAM
i spent ten minutes waiting for the computer to be unfrozen + write up these comments. i'm already being generous enough with my time after those 10 minutes, especially by describing how i tested in different browsers. if you can't see that, not my problem!
I'm with the commenter: when my first (or within 15 minute) impression of a vendor or product is craptastic I cut my losses and consider myself lucky for dodging a longer term bullet.
The commenter was actually very considerate and raised a warning where it might be seen. And they were kind enough to attempt this with two different browsers. After that you can buy my troubleshooting time at its usual hourly rate.
(Because it's a frequent enough issue: I wouldn't see that warning as being about a one-off obscure bug that will affect few people and doesn't matter. It's a warning that the web site probably did not enough take compatibility in consideration - and was approved without such consideration.)
"Runs fine for me" is an absurd bar for reliability / compatibility, no?
Very cool! But I think we can help you improve your video-streaming and teleop (it seems to be pretty low frame rate in the demo). We've built the probably best remote teleop solution for robotics in the market today and it can be embedded anywhere (white-label). Want to get in touch to discuss? You can find my linkedin in my profile.
the video stream will get a lot of improvement for sure!
the teleop is pretty good already i believe, but happy to chat further. feel free to dm / join discord
with congestion control, packet loss mitigation, hardware accelerated encoding, multiple streams, teleop commands on the same channel (required for safety)? Do you host your own TURN server? My point is that robotics companies should focus and push the envelop on the application of their product, not reinvent the wheel on infrastructure and tooling.
I assume you mean Orin Nano? (Jetson nano is EOL and only supports Ubuntu 18). Yes, we have users that run on Orin Nano and it's plug-and-play on any hardware. However note that the Orin Nano doesn't have hardware encoders, so it will take CPU cycles to encode the video, making it a less then ideal choice for teleoperated robots. Cheaper boards like the Orange Pi 5 are a better fit.
All the webrtc-features really only become relevant when you want to control the robot over the Internet, i.e., not just locally where you can assume reliable network.
That's a cool little robot, and I see the appeal, definitely as an educational toy. I don't think the fidelity is there for real autonomous research. But the bigger issue imo is that there's no way this should cost $2000.
You've got a $250 computer, some lidar+camera sensor for maybe $1-200, 6 servos, and cheap plastic. Plus you want to charge a $50/mo software subscription fee for some software product, whatever I guess that's beside the point.
No shade on the idea because low-cost robotics is an unsolved need for the future. But this current iteration is just not competing well with other alternatives. Perhaps this is more of a comment on what we can accomplish in the West vs what's possible in Asia.
Why would I not go for this guy for $1600, and attach an arm? https://www.unitree.com/go2
It's not an apples-to-apples product comparison, but you get the point. There's just so much more raw value offered per dollar elsewhere.
Actually the BOM cost required to make something stable that can execute manipulation tasks well enough is around $1k+ hence our price. You will find very cheap robots that can pretend to do what this one can, but in practice won't work well enough.
As for the unitree robot, this one is not unlocked for development, does not have onboard GPU, and does not have an arm. If you want it, check the price they give, it's very prohibitive.
You could attach a cheap arm to it but it would also not be stable enough for AI algorithms to run it. We're researchers ourselves, we would have made it cheaper if we could, but then you just can't do anything with it.
Our platform will deliver the experience of a real AI robot, anything cheaper than that is kind of a lie - or forces you to assemble and calibrate, which we do for you here. It is just the nature of trying to deliver a really complete product that works, and we want to stand for that.
EDIT: You can take a look at our autonomous demos there, you need something reliable for these: https://docs.innate.bot/welcome/mars-example-use-cases
As a side note, the previous generation of research platforms for that size made in Asia were the Turtlebots, which go for that same price, but without GPU, arm...
I would say the problem is that most manufacturers, including chinese, sell you platforms that are not reliable enough for AI manipulation, and there's a race to the bottom for it, to which we try not to participate to
Coming with a Lidar out of the box seems nice.
Does the MARS hardware really remove the hidden extras (computer with a gpu) mentioned as the downside of HF SO-101 or LeKiwi? While a jetson is good for inference, I feel like to train VLAs you would need access to a powerful GPU regardless. For Lerobot based hardware training ACT was relatively low profile if you use low resolution for the camera feeds, but with increased resolution or with more than one camera I already saw needing more than 8GB of VRAM. If VLA is on the table, finetuning something like the open sourced version of pi0 should already necessitate access to more than one 4090 or above I think.
Also, do you have plans for community-level datasets? I think Lerobot sort of does this with their data recording pipeline and HF integration.
One of our objectives here was to fix everything that we don't like about the SO-101 and the Kiwi, which have several hardware and software flaws in our view. Including, yes, the constant need for a computer to simply run your robot.
The training does require external GPUs (but we provide that infra for free, straight from the app!), but the onboard jetson can run models trained though, as you can see in the examples. Everything you see in the vids is running onboard when it comes to manipulation, because we use a special version of ACT made specifically by us for this robot, that also includes a reward model (like DYNA does).
We have developed this system to also be able to run the other components smoothly so it also does SLAM, and has room for more processing even when running our ACT.
Now indeed this cannot run Pi-0 but from our experience - and the whole community in general - VLAs are not particularly better than ACT in the low data regime, and need a lot more compute.
As for community-level datasets, yes this is the plan. Anything you train can already be shared with others - just share the files. We didn't develop a centralized place for sharing datasets and behaviors but it is on the plan.
Looks awesome!
This isn't so clear though: https://docs.innate.bot/main/software/basic/connecting-to-ba...
> BASIC is accessible for free to all users of Innate robots for 300 cumulative hours - and probably more if you ask us.
Is BASIC used just to create the behaviours or to run them too? It sounds like this is an API you host that turns a behaviour like "pick up socks" into ROS2 motor commands for the robot. Are you open sourcing this too, so anyone can run the (presumably GPU heavy) backend?
Does the robot needs an internet connection to work?
Also, more importantly, what does it look like with googly eyes stuck on?
On BASIC: Yes it does require an internet connection and until we figure out how this works for you it will remain free of use!
It is required to run them, not to create them. And it's not about running "pick_up_socks", this one can already run on your robot. BASIC is required to chain it with other tasks such as navigating to another part of your house and then running another skill to drop the sock somewhere for example
Thank you for the remark, we will make it clearer in the docs
As a consequence: The robot does not necessarily require Internet to run, but if you want it to chain tasks while talking and using memory, yes it does.
As for the googly eyes, give me a minute...
I'll buy one...but your discount code doesn't work. It says "Enter a valid discount code".
Fixed it! It was only possible to use a given number of times
What motors do you use for the arm and what interfaces do you provide (position, velocity, effort)? How long does the battery last when idling?
On the battery, at least 5 hours when idle, and 3 when moving at least
On the motors, these are dynamixels from robotis, and we provide all three of position, velocity, and effort on the low-level SDK (in ROS2 too)
we also provide interfaces for IK
and much higher interfaces for interaction and ai manipulation. Like directly recording episodes of training data so that the arm can use a VLA instead of simple IK.
> our novel SDK called BASIC
Sigh.
yeah it's a tribute
Doesn’t that name make it confusing for your users? A tribute maybe shouldn’t be identical?
this website froze my computer for several minutes, like i'm browsing the web in 2004. making the impression that the robot will work the same way, very uninterested now
you're the first person to mention that to me and it's very very helpful actually. can i contact you to see what you see?
ain't got time for that. happened in both firefox (locked down with lot of privacy settings/adblockers) and chrome (no privacy/adblockers etc). froze entire computer on both browsers for a few minutes for each browser. my computer has like 16gb of RAM
Runs fine for me. Perhaps a bit more investigation on your end before putting fault on the site and authors?
i spent ten minutes waiting for the computer to be unfrozen + write up these comments. i'm already being generous enough with my time after those 10 minutes, especially by describing how i tested in different browsers. if you can't see that, not my problem!
I'm with the commenter: when my first (or within 15 minute) impression of a vendor or product is craptastic I cut my losses and consider myself lucky for dodging a longer term bullet.
The commenter was actually very considerate and raised a warning where it might be seen. And they were kind enough to attempt this with two different browsers. After that you can buy my troubleshooting time at its usual hourly rate.
(Because it's a frequent enough issue: I wouldn't see that warning as being about a one-off obscure bug that will affect few people and doesn't matter. It's a warning that the web site probably did not enough take compatibility in consideration - and was approved without such consideration.)
"Runs fine for me" is an absurd bar for reliability / compatibility, no?
Very cool! But I think we can help you improve your video-streaming and teleop (it seems to be pretty low frame rate in the demo). We've built the probably best remote teleop solution for robotics in the market today and it can be embedded anywhere (white-label). Want to get in touch to discuss? You can find my linkedin in my profile.
the video stream will get a lot of improvement for sure! the teleop is pretty good already i believe, but happy to chat further. feel free to dm / join discord
Don't underestimate the challenges of making remote-teleop work reliably and efficiently with low-latency via the Internet: https://transitiverobotics.com/blog/streaming-video-from-rob....
webrtc for video stream is what we have in the pipeline for improving the stream yep!
with congestion control, packet loss mitigation, hardware accelerated encoding, multiple streams, teleop commands on the same channel (required for safety)? Do you host your own TURN server? My point is that robotics companies should focus and push the envelop on the application of their product, not reinvent the wheel on infrastructure and tooling.
i don't know if it is that necessary for just phone control, wdyt?
do you have anything plug-and-play for jetson nano?
I assume you mean Orin Nano? (Jetson nano is EOL and only supports Ubuntu 18). Yes, we have users that run on Orin Nano and it's plug-and-play on any hardware. However note that the Orin Nano doesn't have hardware encoders, so it will take CPU cycles to encode the video, making it a less then ideal choice for teleoperated robots. Cheaper boards like the Orange Pi 5 are a better fit.
All the webrtc-features really only become relevant when you want to control the robot over the Internet, i.e., not just locally where you can assume reliable network.