What if Tesla Bot / Optimus actually becomes a big deal success in the near future (<6 years?) Up until recently I would be quite surprised, but after further reflection now I’m not so sure.
Here’s my best “bull case:”
Boston Dynamics and things like this https://www.youtube.com/watch?v=zXbb6KQ0xV8 establish that getting robots to walk around over difficult terrain is possible with today’s tech, it just takes a lot of engineering talent and effort.
So Tesla will probably succeed, within a few years, at building a humanoid robot that can walk around and pick up some kinds of objects and open doors and stuff like that. All the stuff that Boston Dynamic’s spot can do. Motors will be weak and battery life will be short though.
What about price? Thanks to economies of scale, if Tesla is producing 100,000+ bots a year the price should be less than the price of a Tesla car.
The hard part is getting it to actually do useful tasks, to justify producing that many (or indeed to justify producing it at all). Just walking around and opening doors is already useful (see: Boston Dynamics being able to sell a few thousand Spot robots) but it seems like a pretty small niche.
As Tesla’s adventures with FSD beta show, navigating the real world is hard (at least if you have a brain the size of an ant.) However if you look at it differently, the lesson is maybe the opposite: Navigating the real world is easy so long as you have lots of training data and are willing to accept occasional crashes. (Teslas have, like, one plausibly-necessary intervention every half hour of city driving now, according to my brother.) If a Tesla Bot falls over, or bumps into a wall, or drops what it is carrying about once every half hour… that’s actually fine? Not a big deal if it’s job is to stock shelves or sweep floors or flip burgers. Replace half your employees with bots & (since they can work 24⁄7) have them do the jobs of all the employees, and then have your remaining employees job be micromanaging the bots and cleaning up the messes they constantly make. Recoup your investment in less than a year.
And (the bullish hope continues) you can use a much more aggressive, risk-accepting training strategy if you aren’t trying to completely avoid failures. Whereas Tesla is trying to create an AI driver without letting it kill anyone along the way, for their Optimus bots it’s totally fine if they are constantly falling down and dropping stuff for a period until they get good. So perhaps the true rate of dropping objects and falling over could be fairly easily driven down to once a day or less—comparable to clumsy human workers.
DeepMind has done a few experiments where AIs in a simulated house hear english-language instructions like “go to the bathroom put all the green stuff on the table” and “follow me.” It works. Now, that’s just a simulated house rather than a real house. But maybe an AI pre-trained in simulated environments can then be retrained/fine-tuned in real houses with real humans giving instructions. No doubt it would take many many hours of trial and error to get good, but… Tesla can easily afford that. They can build 100,000 robots and probably have 100,000 beta testers paying them for the privilege of providing training data. If the project totally flops? They’ve lost maybe five billion dollars or so. Not the end of the world for them, it’s not like they’d be betting the company on it. And they’ve been willing to bet the company on things in the not-so-distant past.
Yes, this is massively more economically useful than any similar robot ever made. It would be regarded as a huge leap forward in AI and robotics. But maybe this isn’t strong reason to be skeptical—after all, no robot has ever had 100,000 instances to learn from before except Tesla’s self-driving cars, and those actually do pretty well if you look at it in the right way. (Am I forgetting something? Has a deep learning experiment with 100,000 robots doing tasks in the real world been done before?)
So what’s the bearish case? Or counterarguments to the bullish case?
Stocking shelves maybe? This seems like the best answer so far.
Packing boxes in warehouses (if that’s not already done by robots?)
What about flipping burgers?
What about driving trucks? Specifically, get an already-autonomous truck and then put one of these bots in it, so that if you need some physical hands to help unload the truck or fill it up with gas or do any of those ordinary menial tasks associated with driving long distances, you’ve got them. (A human can teleoperate remotely when the need arises)
Maybe ordinary factory automation? To my surprise regular factory robots cost something like $50,000; if that’s because they aren’t mass-produced enough to benefit from economies of scale, then Tesla can swoop in with $20,000 humanoid robots and steal market share. (Though also the fact that regular factory robots cost so much is evidence that Tesla won’t be able to get the price of their bots down so low) https://insights.globalspec.com/article/4788/what-is-the-real-cost-of-an-industrial-robot-arm
Maybe cleaning? In theory a robot like this could handle a mop, a broom, a dust wand, a sponge, a vacuum, etc. Could e.g. take all the objects off your sink, spray it with cleaner and wipe it down, then put all the objects back.
Maybe cooking? Can chop carrots and stuff like that. Can probably follow a recipe, albeit hard-coded ones. If it can clean, then it can clean up its own messes.
I wish I had a better understanding of the economy so I could have more creative ideas for bot-ready jobs. I bet there are a bunch I haven’t thought of.
...
Tesla FSD currently runs on hopium: People pay for it and provide training data for it, in the hopes that in a few years it’ll be the long-prophecied robocar.
Maybe a similar business model could work for Optimus. If they are steadily improving it and developing an exponentially growing list of skills for it, people will believe that in a few years it’ll be a fully functioning household servant, and then people will buy it and provide training data even if all it can do is clumsily move around the house and put small objects from the floor to the shelves and even if it needs to be picked up off the floor multiple times a day.
What if Tesla Bot / Optimus actually becomes a big deal success in the near future (<6 years?) Up until recently I would be quite surprised, but after further reflection now I’m not so sure.
Here’s my best “bull case:”
Boston Dynamics and things like this https://www.youtube.com/watch?v=zXbb6KQ0xV8 establish that getting robots to walk around over difficult terrain is possible with today’s tech, it just takes a lot of engineering talent and effort.
So Tesla will probably succeed, within a few years, at building a humanoid robot that can walk around and pick up some kinds of objects and open doors and stuff like that. All the stuff that Boston Dynamic’s spot can do. Motors will be weak and battery life will be short though.
What about price? Thanks to economies of scale, if Tesla is producing 100,000+ bots a year the price should be less than the price of a Tesla car.
The hard part is getting it to actually do useful tasks, to justify producing that many (or indeed to justify producing it at all). Just walking around and opening doors is already useful (see: Boston Dynamics being able to sell a few thousand Spot robots) but it seems like a pretty small niche.
As Tesla’s adventures with FSD beta show, navigating the real world is hard (at least if you have a brain the size of an ant.) However if you look at it differently, the lesson is maybe the opposite: Navigating the real world is easy so long as you have lots of training data and are willing to accept occasional crashes. (Teslas have, like, one plausibly-necessary intervention every half hour of city driving now, according to my brother.) If a Tesla Bot falls over, or bumps into a wall, or drops what it is carrying about once every half hour… that’s actually fine? Not a big deal if it’s job is to stock shelves or sweep floors or flip burgers. Replace half your employees with bots & (since they can work 24⁄7) have them do the jobs of all the employees, and then have your remaining employees job be micromanaging the bots and cleaning up the messes they constantly make. Recoup your investment in less than a year.
And (the bullish hope continues) you can use a much more aggressive, risk-accepting training strategy if you aren’t trying to completely avoid failures. Whereas Tesla is trying to create an AI driver without letting it kill anyone along the way, for their Optimus bots it’s totally fine if they are constantly falling down and dropping stuff for a period until they get good. So perhaps the true rate of dropping objects and falling over could be fairly easily driven down to once a day or less—comparable to clumsy human workers.
DeepMind has done a few experiments where AIs in a simulated house hear english-language instructions like “go to the bathroom put all the green stuff on the table” and “follow me.” It works. Now, that’s just a simulated house rather than a real house. But maybe an AI pre-trained in simulated environments can then be retrained/fine-tuned in real houses with real humans giving instructions. No doubt it would take many many hours of trial and error to get good, but… Tesla can easily afford that. They can build 100,000 robots and probably have 100,000 beta testers paying them for the privilege of providing training data. If the project totally flops? They’ve lost maybe five billion dollars or so. Not the end of the world for them, it’s not like they’d be betting the company on it. And they’ve been willing to bet the company on things in the not-so-distant past.
Yes, this is massively more economically useful than any similar robot ever made. It would be regarded as a huge leap forward in AI and robotics. But maybe this isn’t strong reason to be skeptical—after all, no robot has ever had 100,000 instances to learn from before except Tesla’s self-driving cars, and those actually do pretty well if you look at it in the right way. (Am I forgetting something? Has a deep learning experiment with 100,000 robots doing tasks in the real world been done before?)
So what’s the bearish case? Or counterarguments to the bullish case?
I think the problem here is clear use cases. What is the killer app for the minimum viable robot?
Yeah, that’s the crux...
Stocking shelves maybe? This seems like the best answer so far.
Packing boxes in warehouses (if that’s not already done by robots?)
What about flipping burgers?
What about driving trucks? Specifically, get an already-autonomous truck and then put one of these bots in it, so that if you need some physical hands to help unload the truck or fill it up with gas or do any of those ordinary menial tasks associated with driving long distances, you’ve got them. (A human can teleoperate remotely when the need arises)
Maybe ordinary factory automation? To my surprise regular factory robots cost something like $50,000; if that’s because they aren’t mass-produced enough to benefit from economies of scale, then Tesla can swoop in with $20,000 humanoid robots and steal market share. (Though also the fact that regular factory robots cost so much is evidence that Tesla won’t be able to get the price of their bots down so low) https://insights.globalspec.com/article/4788/what-is-the-real-cost-of-an-industrial-robot-arm
Maybe cleaning? In theory a robot like this could handle a mop, a broom, a dust wand, a sponge, a vacuum, etc. Could e.g. take all the objects off your sink, spray it with cleaner and wipe it down, then put all the objects back.
Maybe cooking? Can chop carrots and stuff like that. Can probably follow a recipe, albeit hard-coded ones. If it can clean, then it can clean up its own messes.
I wish I had a better understanding of the economy so I could have more creative ideas for bot-ready jobs. I bet there are a bunch I haven’t thought of.
...
Tesla FSD currently runs on hopium: People pay for it and provide training data for it, in the hopes that in a few years it’ll be the long-prophecied robocar.
Maybe a similar business model could work for Optimus. If they are steadily improving it and developing an exponentially growing list of skills for it, people will believe that in a few years it’ll be a fully functioning household servant, and then people will buy it and provide training data even if all it can do is clumsily move around the house and put small objects from the floor to the shelves and even if it needs to be picked up off the floor multiple times a day.