Flipping a pancake is extremely difficult because each pancake is different. I know that these videos must be cherry-picked but to be able to train a Robot to do this just by demonstrating feels like a massive leap.
Solving any one problem with robotic manipulation isn’t all that hard. It takes a lot of trial and error, but in general if the task is constrained you can solve it reliably. The trick is to solve *new* tasks without resorting to all that fine tuning every time. Which is what Russ is claiming here. He’s training an LLM with a corpus of one-off policies for solving specific manipulation tasks, and claiming to get robust ad hoc policies from it for previously unsolved tasks.
If this actually works, it’s pretty important. But that’s the core claim: that he can solve ad hoc tasks without training or hand tuning.
I know it isn’t as open ended as plenty of more important problems in robotics, but this doesn’t strike me as easy at all.
I’ve only dabbled in robotics as an entry level hobbiest, so I really don’t know the answer.
It's hard to point to a single thing that would make "flipping pancakes" intractable, it's sort of the other way around, to usefully flip pancakes in the same way as a person takes a lot of skills chained together.
The "door problem" is a sort of compendium of many real-world skills, identifying the door, understanding its affordances and how to grip / manipulate them, whether to push or pull the door, predicting the trajectory of the door when opened, estimating the mass of the door and applying the right amount of force, understanding if there any springs or pulls on the door and how it must be held to traverse through it. Etc. There are also a ton of things I'm missing that are so fundamental one tends to take them for granted, like knowing your own size and that you can't fit through a tiny doorway.
I think you can ramp towards the "door problem" in difficulty by slowly relaxing constraints. A video linked above (not article) shows "can flip a pancake successfully with a particular pan (you are already holding) and pancake with a fixed camera and visual markers". Ok, now do it in varying lighting conditions. With no visual markers. With different camera views. Different pancakes. Real pancakes (which are not rigid, and sometimes stick to the pan). Different pans. Now you have to pick up the pan. Use a stove. Different stoves. Identify griddle vs pan and use the right flipping technique. Find everything and do it all in a messy kitchen... eventually you're getting to same ballpark as the "door problem".
(I also ignored that door opening is generally done by mobile robots of a certain weight class which tend to be more expensive than a stationary arm with enough strength to pick up a spatula or hold a pan).
There is a steep difficulty gradient from "works in the lab" to "works under semi-controlled real world conditions" to "works in uncontrolled real-world situations".
The way things are going is sensors (cameras, force, etc) and neural networks. You let the robot try a bunch of ways of opening doors, sometimes it doors itself in the face, eventually it'll figure out good places to stand based on what the door looks like. The more doors you make it try to open hopefully the better it gets at generalising over the task of opening doors. The hacks/heuristics are really still there but the robot is supposed to learn them.
> Surely people don’t stop in front of a door and begin planning things out, rather they seem to go for it and adjust on the fly, is this an approach that won’t work in robotics? Why not?
Yeah, figuring out how to do this is basically "the problem". Most people don't have a sense or feeling of "planning things out" as they open a door because we reached "unconscious competence" at that task. We definitely have predictions of what is going to happen as we start opening the door based on prior experience and our observations so far. If reality diverges from our expectations we will experience surprise, make some new predictions, take actions to resolve the surprise, etc.
Not sure that anyone has ever studied how people open doors in detail, it'd be interesting. I bet there are a ton of subtle micro behaviours. One that I know is, if you hear kids running in the house it is a good idea to keep a foot planted in front of you as you approach the door, because those guys will absolutely fling or crash doors open right into your face.
> He’s training an LLM with a corpus of one-off policies for solving specific manipulation tasks, and claiming to get robust ad hoc policies from it for previously unsolved tasks.
In my opinion that really is a good definition of intelligence, and puts this technique at the forefront of machine intelligence.
If instead your choice is your high school bully or a robot, well for now pick the bully. Because that robot isn’t even being vicious and will hurt more.
Rodney Brooks at the MIT AI Lab was a big advocate of something called "series elastic actuators." The idea was was that you didn't allow motors to directly turn robot joints. Instead, all motors acted through some kind of elastic. And the robots could also measure how much resistance they encountered and back off.
MIT had a number of demos of robots that played nicely around fragile humans. I remember video of a grad student stopping a robot arm with their head.
Now, using series elastic actuators will sacrifice some amount of speed or precision. You wouldn't want to do it for industrial robots. And of course, robots also tend to be heavy and made of metal, so even if they move gently, they still pose real risks.
But real progress has been made on these problems.
I'll take the robot, thanks.
So, the moment your system needs this kind of data and that kind of data, oh and btw it needs a few hundreds of thousands of examples of all those kinds of data, that's very clear to me that it's far away from being capable of any kind of generalisation, any kind of learning general behaviour.
So that's "60 difficult, dexterous skills" today, "1,000 by the end of 2024", and when the aliens land to explore the ruins of our civilisation your robot will still be training on the 100,000'th "skill". And still will fall over and die when the aliens greet it.