″ That is true, but the focus remains on what isdifferentthis time. I may be wrong, but I strongly suspect that no one views better robots as different in kind from the sort of automation we saw in the 1980s. This is just an incremental improvement over the kinds of things we could do before. ”
Here’s what I think is different about it this time. Specifically, the 1980s methods involved basically : you identify the task to be automated. You hire an automation engineering firm to do the task. This would involve custom “fixtures”, custom mechanical assemblies, and very carefully designed assembly lines. Then a team of programmers has to very carefully put together a sequence of commands to complete the task. If you binge-watch “how it’s made” videos, you will see hundreds of examples of this. Cake factories where a dipping machine is custom made for the shape of the cakes going into the chocolate dip, where it can’t make a different kinds of cake or change the chocolate ratio or make rice crispies instead or learn how to clear faults where the cake has clogged up the machine.
One big flaw with this is if you have a new kind of cake you want to make, you can’t just send a description of your new recipe to the factory and have it make it. Nor can the equipment try different strategies for cake making and self-optimize for higher speed and less errors. Nor can the equipment be given a new set of robot hardware that has slightly different performance parameters and self-adjust to make the cakes using these different robot actuators. Nor can it avoid “hurting itself” by carefully planning each motion and making sure planned motions are not going to hit anything in the environment. Some human has to set all this up. Some human has to manually adjust timings, to unclog it whenever it faults, and so on.
Today’s methods show all of the above is possible in the immediate future. Also, we can potentially build massive frameworks, where automating even small tasks is easy because the framework is stable, easy to use, easy to connect to compatible hardware. And you just go on the “app store” for the framework and rent the pieces you need to do a task. Go rent a classifier that can recognize most objects seen through a camera. A physics modeler that is both learning and already pre-trained for most ordrinary objects in a factory. Scoring software that can measure outcomes. Just grab all the pieces and put together your automation “app”.
And obviously we can go to either waldos that can do many tasks and don’t need custom mechanical fixtures, or we find a way to rapid-design fixtures and get them installed and working automatically.
And, hopefully, robots making cakes will be able to share knowledge back to the cloud, so that other robots elsewhere making transmissions get slightly better, or if an employee drops a cake on the factory floor, they know what it is.
So ultimately, while it’ll take a long time to actually build all this, ultimately we can automate all retail stores, all warehouses, all farms, and all factories. Billions of jobs. I don’t know how to automate service jobs like cutting hair with the current state of the art, but even if we can’t, if half the population of most countries are out of work and trying to re-train to cut hair or something, it crashes the labor market for barbers, etc.
″ That is true, but the focus remains on what is different this time. I may be wrong, but I strongly suspect that no one views better robots as different in kind from the sort of automation we saw in the 1980s. This is just an incremental improvement over the kinds of things we could do before. ”
Here’s what I think is different about it this time. Specifically, the 1980s methods involved basically : you identify the task to be automated. You hire an automation engineering firm to do the task. This would involve custom “fixtures”, custom mechanical assemblies, and very carefully designed assembly lines. Then a team of programmers has to very carefully put together a sequence of commands to complete the task. If you binge-watch “how it’s made” videos, you will see hundreds of examples of this. Cake factories where a dipping machine is custom made for the shape of the cakes going into the chocolate dip, where it can’t make a different kinds of cake or change the chocolate ratio or make rice crispies instead or learn how to clear faults where the cake has clogged up the machine.
One big flaw with this is if you have a new kind of cake you want to make, you can’t just send a description of your new recipe to the factory and have it make it. Nor can the equipment try different strategies for cake making and self-optimize for higher speed and less errors. Nor can the equipment be given a new set of robot hardware that has slightly different performance parameters and self-adjust to make the cakes using these different robot actuators. Nor can it avoid “hurting itself” by carefully planning each motion and making sure planned motions are not going to hit anything in the environment. Some human has to set all this up. Some human has to manually adjust timings, to unclog it whenever it faults, and so on.
Today’s methods show all of the above is possible in the immediate future. Also, we can potentially build massive frameworks, where automating even small tasks is easy because the framework is stable, easy to use, easy to connect to compatible hardware. And you just go on the “app store” for the framework and rent the pieces you need to do a task. Go rent a classifier that can recognize most objects seen through a camera. A physics modeler that is both learning and already pre-trained for most ordrinary objects in a factory. Scoring software that can measure outcomes. Just grab all the pieces and put together your automation “app”.
And obviously we can go to either waldos that can do many tasks and don’t need custom mechanical fixtures, or we find a way to rapid-design fixtures and get them installed and working automatically.
And, hopefully, robots making cakes will be able to share knowledge back to the cloud, so that other robots elsewhere making transmissions get slightly better, or if an employee drops a cake on the factory floor, they know what it is.
So ultimately, while it’ll take a long time to actually build all this, ultimately we can automate all retail stores, all warehouses, all farms, and all factories. Billions of jobs. I don’t know how to automate service jobs like cutting hair with the current state of the art, but even if we can’t, if half the population of most countries are out of work and trying to re-train to cut hair or something, it crashes the labor market for barbers, etc.