The challenge is that the current approaches that are very novel broadly have proven relatively less useful for things like robotic manual labor.
The crux of the current AI craze is machine learning, which needs all the sensory information the human perceives that is relevant to the task and all of the things the human does in response in every detail. You can’t tell a model to “just feel out how that screwdriver is going and make sure you don’t strip the head or cross-thread it by mistake”, it has no abstraction, it can only derive the most likely inputs from training data based on the sensory information.
This works reasonably well for vehicle operation, where the human generally only has sight and hearing, and generally only operates a few relatively simple controls, wheels, pedals, knobs. Even with that advantage, it needs a mountain of sample data. Millions of miles of drive data to work with and the result is still a bit janky, kind of like a human with about 3000 miles of total lifetime driving experience.
Once you want to extend this approach to anything involving a more comprehensive use of hand dexterity and everything you can feel, then it becomes a problem. If you can imagine yourself tele-operating a robot with VR headset and controllers to do a task, then they have a path to get training data. That’s how a lot of these ‘chore-bot’ ambitions are trying to be realized, workers tele-operating for demos and training data, in hopes of getting enough training data to do things competently one day. But imagine trying to use a manual screw driver in that tele-operation scenario, making sure it’s lined up, doesn’t cross thread, and you don’t strip the head. I couldn’t imagine doing something even that simple in the real world without just being able to feel the screwdriver in my hand, to feel how the screw is going in or not going in, if the resistance feels right and consistent. If it’s a situation that makes sense to use a driver, then probably could tele-operate it, but in cases where we need the finesse of a manual tool, those are going to be super challenging to address. Even using a driver, I always manually start it, either with my fingers or a screwdriver, and the driver comes in when alignment is well assured.
Now the long standing approaches are of course already a big factor in automating construction. The new approaches can also make it easier to take a task that was almost within the reach of robotic choreogrpahy but foiled by something like unexpected weight and wind, so the relatively newer AI approaches help make it easy to correct for deviations to provide a steady environment to execute the programmed movements. But the versalitility of a person with full use of their hands and feelings is going to be a tough road unless some breakthrough happens.
The challenge is that the current approaches that are very novel broadly have proven relatively less useful for things like robotic manual labor.
The crux of the current AI craze is machine learning, which needs all the sensory information the human perceives that is relevant to the task and all of the things the human does in response in every detail. You can’t tell a model to “just feel out how that screwdriver is going and make sure you don’t strip the head or cross-thread it by mistake”, it has no abstraction, it can only derive the most likely inputs from training data based on the sensory information.
This works reasonably well for vehicle operation, where the human generally only has sight and hearing, and generally only operates a few relatively simple controls, wheels, pedals, knobs. Even with that advantage, it needs a mountain of sample data. Millions of miles of drive data to work with and the result is still a bit janky, kind of like a human with about 3000 miles of total lifetime driving experience.
Once you want to extend this approach to anything involving a more comprehensive use of hand dexterity and everything you can feel, then it becomes a problem. If you can imagine yourself tele-operating a robot with VR headset and controllers to do a task, then they have a path to get training data. That’s how a lot of these ‘chore-bot’ ambitions are trying to be realized, workers tele-operating for demos and training data, in hopes of getting enough training data to do things competently one day. But imagine trying to use a manual screw driver in that tele-operation scenario, making sure it’s lined up, doesn’t cross thread, and you don’t strip the head. I couldn’t imagine doing something even that simple in the real world without just being able to feel the screwdriver in my hand, to feel how the screw is going in or not going in, if the resistance feels right and consistent. If it’s a situation that makes sense to use a driver, then probably could tele-operate it, but in cases where we need the finesse of a manual tool, those are going to be super challenging to address. Even using a driver, I always manually start it, either with my fingers or a screwdriver, and the driver comes in when alignment is well assured.
Now the long standing approaches are of course already a big factor in automating construction. The new approaches can also make it easier to take a task that was almost within the reach of robotic choreogrpahy but foiled by something like unexpected weight and wind, so the relatively newer AI approaches help make it easy to correct for deviations to provide a steady environment to execute the programmed movements. But the versalitility of a person with full use of their hands and feelings is going to be a tough road unless some breakthrough happens.