Any machine tool worth a damned can rip your arm off and/or kill you. Never google degloving.
The same is true for robots. Safety where they have to share space with humans is distinctly non-trivial.
This matters far more than the compute or feedback systems required. We have robots that do laundry and wash dishes, they shut off when you open them, or sometimes actually lock while in use so you can't open them, for good reasons.
Have you seen a Costco Food Court? The first time I saw it I thought of how inefficient it was. All these people crowded around a pickup window waiting for their number to be called.
They could easily build some sort of notification system so people would get a message on their phone when their order was ready. They could be shopping around the store rather than all waiting here.
Such an obvious opportunity for automation.
But after watching how they actually work and reflecting on it for a bit, I started to see why they don't do that.
- Letting people walk away likely means more orders will be piling up in their limited working space waiting to be picked up.
- These workers were fast. Juggling multiple orders at the same time. The "state" of the order is all in one place, a physical receipt and that state doesn't really change. Adding software to their workflow would almost certainly slow it all down (for reference, see hospital software).
- You'd have to train people on how to use the notification system. What happens when the system changes? What about security? What if someone were to hijack it and start sending out fake notifications?
My takeaway here is that every decision has trade-offs so before you think something can be automated you should understand what the manual process is first and why it works the way it does.
Most environments are not robot friendly I think. I think in the future people will build robot friendly X (like home) to ease the software/hardware requirements. Actually governments might build those buildings for UBC programs.
A lot of it is general complexity. A lot of situations require not only complex behaviour that's quite tricky to program into a robot, but the ability to deal with situations outside of their control. Like on a production line, that's not an issue. There's no weather, there are no humans or other animals in the nearby environment, the parts all get transported at the same pace and in the same state, etc.
The robot can basically run a macro that goes from the same few motions over and over again.
Now think about a self-driving car. They're a lot closer to being practical now, but they still have a lot of challenges to deal with that the robot arms in factories don't.
They have to worry about legal issues if they make a mistake. There are other vehicles on the road, including both human driven ones and ones by other companies. There might be pedestrians nearby, or wildlife. Weather has an effect on how they drive, since even a machine can't react as well to flooded or icy roads as normal ones. There may be natural disasters, or ones caused by human/robot mistakes. The programming for this sort of system can't just follow a simple routine, it has to accounts for dozens or even hundreds of possibilities.
Same situation in many other places robots could be used in. Shops and restaurants, construction sites, on the street doing law enforcement and traffic work, etc. Unless you have an incredibly well-designed, incredibly flexible machine that can deal with unforeseen circumstances, it may be cheaper, easier and less legally risky to have a human do the work.
The easiest solution to this is to retool the layout of the space the robot will occupy to fit the needs of said robot. So, a shop or restaurant would be designed in a way that the automated setup would be closer to an oven or dishwasher and the process wouldn't interact with humans beyond the ordering and receiving aspects.
But that's expensive and time consuming to do, and in a lot of cases may hurt the business needs of the company. I don't think a lot of people want to get their food from the equivalent of a vending machine at a high-end restaurant, or have it delivered by a glorified roomba.
Unless humans become extremely expensive to hire and robotics becomes extremely cheap, that probably won't be worth the hassle to change.
Its economics, I'm not going to pay $100k for a robot to empty my dishwasher. I won't even pay $1k to purchase a robot vacuum.
There's no ROI for it at this point in time for consumers. Robots have been automating various things in factories for decades, its not new. Making the robots look like humans and act like humans doesn't really provide value does it?
For example how is a $100-200k robot cashier better than self-checkout or having a teenager get paid $12/hour?
I get that economics plays a huge role, but as with almost every technology, it starts off super expensive and is only used by big corporations (factories, in this case). Then it gets cheaper, and disruptive technologies appear that knock down the incumbents. A surge in demand follows, especially when the technology is actually useful and makes life easier. The history of PCs is a good example.
The process looks good for code and maths, but code has a fast feedback loop starting with "does it compile" and "do these unit tests pass?". It's OK to train models which start off with compiler errors 99.99% of the time and keep improving until it's >0.1% and it can just read and correct when those remaining errors happen.
Do that with a CNC milling machine, and it will drill through its own robotic arms a billion times before it gets good.
Remember: Tesla's fleet has about a trillion miles of real-world experience at this point, and yet the cybercab is extremely geofenced.
In my opinion the hardware isn't there yet for general purpose robotics and the cost of training the AIs to do the work is still too high. Maybe if the AI bubble pops we'll see GPUs going for cheap so robotic AI training will accelerate.
Any machine tool worth a damned can rip your arm off and/or kill you. Never google degloving.
The same is true for robots. Safety where they have to share space with humans is distinctly non-trivial.
This matters far more than the compute or feedback systems required. We have robots that do laundry and wash dishes, they shut off when you open them, or sometimes actually lock while in use so you can't open them, for good reasons.
I hope this is not a bad analogy but here goes.
Have you seen a Costco Food Court? The first time I saw it I thought of how inefficient it was. All these people crowded around a pickup window waiting for their number to be called.
They could easily build some sort of notification system so people would get a message on their phone when their order was ready. They could be shopping around the store rather than all waiting here.
Such an obvious opportunity for automation.
But after watching how they actually work and reflecting on it for a bit, I started to see why they don't do that.
- Letting people walk away likely means more orders will be piling up in their limited working space waiting to be picked up.
- These workers were fast. Juggling multiple orders at the same time. The "state" of the order is all in one place, a physical receipt and that state doesn't really change. Adding software to their workflow would almost certainly slow it all down (for reference, see hospital software).
- You'd have to train people on how to use the notification system. What happens when the system changes? What about security? What if someone were to hijack it and start sending out fake notifications?
My takeaway here is that every decision has trade-offs so before you think something can be automated you should understand what the manual process is first and why it works the way it does.
I agree not every aspect of business is worth automating.
Most environments are not robot friendly I think. I think in the future people will build robot friendly X (like home) to ease the software/hardware requirements. Actually governments might build those buildings for UBC programs.
A lot of it is general complexity. A lot of situations require not only complex behaviour that's quite tricky to program into a robot, but the ability to deal with situations outside of their control. Like on a production line, that's not an issue. There's no weather, there are no humans or other animals in the nearby environment, the parts all get transported at the same pace and in the same state, etc.
The robot can basically run a macro that goes from the same few motions over and over again.
Now think about a self-driving car. They're a lot closer to being practical now, but they still have a lot of challenges to deal with that the robot arms in factories don't.
They have to worry about legal issues if they make a mistake. There are other vehicles on the road, including both human driven ones and ones by other companies. There might be pedestrians nearby, or wildlife. Weather has an effect on how they drive, since even a machine can't react as well to flooded or icy roads as normal ones. There may be natural disasters, or ones caused by human/robot mistakes. The programming for this sort of system can't just follow a simple routine, it has to accounts for dozens or even hundreds of possibilities.
Same situation in many other places robots could be used in. Shops and restaurants, construction sites, on the street doing law enforcement and traffic work, etc. Unless you have an incredibly well-designed, incredibly flexible machine that can deal with unforeseen circumstances, it may be cheaper, easier and less legally risky to have a human do the work.
The easiest solution to this is to retool the layout of the space the robot will occupy to fit the needs of said robot. So, a shop or restaurant would be designed in a way that the automated setup would be closer to an oven or dishwasher and the process wouldn't interact with humans beyond the ordering and receiving aspects.
But that's expensive and time consuming to do, and in a lot of cases may hurt the business needs of the company. I don't think a lot of people want to get their food from the equivalent of a vending machine at a high-end restaurant, or have it delivered by a glorified roomba.
Unless humans become extremely expensive to hire and robotics becomes extremely cheap, that probably won't be worth the hassle to change.
Its economics, I'm not going to pay $100k for a robot to empty my dishwasher. I won't even pay $1k to purchase a robot vacuum.
There's no ROI for it at this point in time for consumers. Robots have been automating various things in factories for decades, its not new. Making the robots look like humans and act like humans doesn't really provide value does it?
For example how is a $100-200k robot cashier better than self-checkout or having a teenager get paid $12/hour?
I get that economics plays a huge role, but as with almost every technology, it starts off super expensive and is only used by big corporations (factories, in this case). Then it gets cheaper, and disruptive technologies appear that knock down the incumbents. A surge in demand follows, especially when the technology is actually useful and makes life easier. The history of PCs is a good example.
It's machine learning.
The process looks good for code and maths, but code has a fast feedback loop starting with "does it compile" and "do these unit tests pass?". It's OK to train models which start off with compiler errors 99.99% of the time and keep improving until it's >0.1% and it can just read and correct when those remaining errors happen.
Do that with a CNC milling machine, and it will drill through its own robotic arms a billion times before it gets good.
Remember: Tesla's fleet has about a trillion miles of real-world experience at this point, and yet the cybercab is extremely geofenced.
First think about the practical, legal and PR challenges of putting self-driving vehicles on public roads.
Putting autonomous robots in amongst humans is that problem, on steroids.
Edit: there’s a reason why industrial robots all have a hazard zone marked around them, and some even have transparent safety screens or cages
First of all, it's extremely expensive. It's cheaper to use biological robots (us). And automation simply isn't feasible in every field.
I think it's both, but the bigger challenge is reliability in unstructured environments.
In my opinion the hardware isn't there yet for general purpose robotics and the cost of training the AIs to do the work is still too high. Maybe if the AI bubble pops we'll see GPUs going for cheap so robotic AI training will accelerate.
Even if the hardware technology is there, the hardware volume isn't, and the hardware factories aren't there to produce the volume.