I have a very opinionated opposing view of this research. A lot of research in this direction is working on raising the floor. Basically they just want the robot to handle a large variety of simple tasks and environments. Fair enough most industrial robots can't handle the smallest changes. But many times they implicitly make the assumption that raising the floor will also raise the ceiling. They assume, if it can generalize at 90%, it might also be able to do far more dextrous tasks that humans can. I think this is completely false, at best if we could do dextrous tasks in 1 environment this can transfer them to other environments with a presumably lower efficiency.
On the other hand, I think a more promising direction is to raise the ceiling of robot arm manipulation sky high. OpenAI kind of did this with Dactyl but I would like to see more of it. Can we get robotic arms to tie a shoelace, knit, perform pottery etc (with an arm like morphology, no special mechanisms). I think this can actually then lead to large scale generalization, kind of what we are seeing happening with NERF's now. I would like a robot arm NERF, overfit to one hard task but reproduces it with human like precision and dexterity. Deepminds approach to me (with GATO, robocat) seems like a red herring, they will never reach the kind of results we want from our arms.
Deepmind is google, and google suffers from chronical dabbling and never shipping products, I'd be surprised if they even care much about generalizing to useful tasks.
And about floor vs ceiling, what's really important is robustness, only robust robots can be deployed in the wild, at this time Dactyl with all its fingers is still too difficult to control, RoboCat got the grippers right, the problem really is they are again doing cute things with large models instead of raising robustness.
I had a more general noobish version of it in my head how all this plays out.
Now that AI based task planning can learn from so few examples and can do extremely general tasks like say the instruction:
"Summarize my gmail twice a day at 7:00 AM and 5:00 PM filtering spam and stuff I don't read" would spawn an agent with a plan to do exactly what you want. And if it could not you could show it once like "recording" a super smart "macro" only this time with AI agents instead of Selenium.
This would be an "inversion" of API way to do things. Pretty soon people are writing tons of these "bots" and multiple cases land in the some(supreme?) court over what is a bot and what is personal property.
I mean I am not blind, there are plenty of no code platforms that work on B2B space. But nothing in the B2C for general purpose domain that do this in a privacy safe manner, at least not one that will shake up the tech companies.
I really really really want to play with robotics at home, but I've discovered it's super expensive to get things like robotic arms at home. Are there robotics platforms that are DIY ready? I really just want to teach a robotic arm to make my coffee in the morning. And I'll apply my life's effort to achieving this only if the cost of the robotics platform is stupid low.
My exact thought as you. I actually have done robotics research in Academia and I find it ridiculous that an arm costs as much as a car when it's just servos in series. I guess economies of scale.
Now that I'm leaving Academia I was trying to figure out how to get a robot arm so that I can continue doing stuff at home. I think an approach like the one here would be good: https://www.trossenrobotics.com/aloha.aspx. A single robot would cost 5k. You could also just buy Dynamixel servos and build it on your own, shouldn't be too hard with 3D printing ~ closer to 3.5 or 4k with that.
You could try to go cheaper by building a dynamixel servo on your own. I would not recommend this, the main difficulty is the controller which requires a lot of tuning, which is what you're paying for when you buy Dynamixel. But the raw parts of a servo are only around 50-100$ I'd guess.
Lastly any arm you build this way is going to be position control which is very different from the human arm that is torque controlled. Trying to get torque control on your arm is all the rage on Youtube these days, perhaps because the motion seems so lifelike. To do that you need to build a Quasi Direct Drive Actuator which is a BLDC servo with a low gear ratio and then try to centralize all the motors to reduce inertia. There is no accepted solution to do this, but the best results come from cable driven mechanisms like AmbiDex. Basically building a human like robot arm is a huge research problem on its own, and if you want to focus on intelligence best avoid this.
The cheapest is to strap things onto a 3 axis 3d printer. They can be had for < $200, which is much less than you could ever build one for.
Beyond that, it's DIY, with a set of strong servos and a few budget microcontrollers/motor controllers, from Alibaba, and...a 3d printer to manufacture the rest, for something like an arm.
Elephant robotics has arms starting around ~$500. I've had good success with their myCobot280-pi. The reach and payload are a bit limited but it's fine for playing around.
* After observing 1000 human-controlled demonstrations, collected in just hours, RoboCat could direct this new arm dexterously enough to pick up gears successfully 86% of the time. *
Interesting, but this doesn’t sound like something that can be useful. For practical industry applications, you need high success rates. Assume you are running operations where dropping/flips/damages to the items that a robot handles are costly and not acceptable.
Its only for a hobbyist pick-and-place, dropping a Lego block or wasting some components is not a big deal.
> Assume you are running operations where dropping/flips/damages to the items that a robot handles are costly and not acceptable.
Such operations are rare and in those cases you're probably not picking them out of a bin to start with. For most items dropping or flipping the object a few times is perfectly acceptable, indeed that's probably how they got into their current position to start with.
Performance on so few examples is impressive, paired with generalizability across broader tasks, and multiple embodiments + environments (and from just visual goals rather than complex verbal instructions) is quite a jump from where we saw Gato at last spring. If representative, seems a strong step toward meaningful autonomous skill acquisition/transference in realistic settings.
"Robots are quickly becoming part of our everyday lives"
Are they really?
This capability will help accelerate robotics research, as it reduces the need for human-supervised training, and is an important step towards creating a general-purpose robot.
The future people are striving to build just seems so fucking creepy to me. I don't really understand the enthusiasm for it. Maybe I would if I saw said future and maybe it would be awesome and I'd regret saying this, but right now, I just don't get it.
Readit News
On the other hand, I think a more promising direction is to raise the ceiling of robot arm manipulation sky high. OpenAI kind of did this with Dactyl but I would like to see more of it. Can we get robotic arms to tie a shoelace, knit, perform pottery etc (with an arm like morphology, no special mechanisms). I think this can actually then lead to large scale generalization, kind of what we are seeing happening with NERF's now. I would like a robot arm NERF, overfit to one hard task but reproduces it with human like precision and dexterity. Deepminds approach to me (with GATO, robocat) seems like a red herring, they will never reach the kind of results we want from our arms.
And about floor vs ceiling, what's really important is robustness, only robust robots can be deployed in the wild, at this time Dactyl with all its fingers is still too difficult to control, RoboCat got the grippers right, the problem really is they are again doing cute things with large models instead of raising robustness.
Now that AI based task planning can learn from so few examples and can do extremely general tasks like say the instruction:
"Summarize my gmail twice a day at 7:00 AM and 5:00 PM filtering spam and stuff I don't read" would spawn an agent with a plan to do exactly what you want. And if it could not you could show it once like "recording" a super smart "macro" only this time with AI agents instead of Selenium.
This would be an "inversion" of API way to do things. Pretty soon people are writing tons of these "bots" and multiple cases land in the some(supreme?) court over what is a bot and what is personal property.
Yes. I know that's absurd. :)
Now that I'm leaving Academia I was trying to figure out how to get a robot arm so that I can continue doing stuff at home. I think an approach like the one here would be good: https://www.trossenrobotics.com/aloha.aspx. A single robot would cost 5k. You could also just buy Dynamixel servos and build it on your own, shouldn't be too hard with 3D printing ~ closer to 3.5 or 4k with that.
You could try to go cheaper by building a dynamixel servo on your own. I would not recommend this, the main difficulty is the controller which requires a lot of tuning, which is what you're paying for when you buy Dynamixel. But the raw parts of a servo are only around 50-100$ I'd guess.
Lastly any arm you build this way is going to be position control which is very different from the human arm that is torque controlled. Trying to get torque control on your arm is all the rage on Youtube these days, perhaps because the motion seems so lifelike. To do that you need to build a Quasi Direct Drive Actuator which is a BLDC servo with a low gear ratio and then try to centralize all the motors to reduce inertia. There is no accepted solution to do this, but the best results come from cable driven mechanisms like AmbiDex. Basically building a human like robot arm is a huge research problem on its own, and if you want to focus on intelligence best avoid this.
Is it really just tuning? That seems like something that would pay off for an aliexpress copycat. A servo is pretty simple/cheap to build, right?
I guess there are only a handful of us who really want high-accuracy, high-torque component servos...
Beyond that, it's DIY, with a set of strong servos and a few budget microcontrollers/motor controllers, from Alibaba, and...a 3d printer to manufacture the rest, for something like an arm.
Don’t think it’s cheap nor a “platform” but maybe it’s enough to get you started ;)
He just recently put out a video on making a carbon fiber arm piece for it.
Interesting, but this doesn’t sound like something that can be useful. For practical industry applications, you need high success rates. Assume you are running operations where dropping/flips/damages to the items that a robot handles are costly and not acceptable.
Its only for a hobbyist pick-and-place, dropping a Lego block or wasting some components is not a big deal.
Alright, let's assume we're not too though, just for completeness sake.
Plenty of industrial applications where an 86% success rate is fine, especially if you can try more than once.
Such operations are rare and in those cases you're probably not picking them out of a bin to start with. For most items dropping or flipping the object a few times is perfectly acceptable, indeed that's probably how they got into their current position to start with.
Are they really?
This capability will help accelerate robotics research, as it reduces the need for human-supervised training, and is an important step towards creating a general-purpose robot.
The future people are striving to build just seems so fucking creepy to me. I don't really understand the enthusiasm for it. Maybe I would if I saw said future and maybe it would be awesome and I'd regret saying this, but right now, I just don't get it.