Autonomous robot people could be constructed right now

In My Humble Opinion
21

However, it seems to me that once a robot could do that, the costs involved would still make it more economical to hire a human than a Robot, one exception would be on places where a Human could be in danger, like in a mine in danger of collapsing, or a nuclear facility after an incident.

Speaking of incidents, experiences with robot makers dealing with a lot of places involved in a disaster, point to the use of robots that are not humanoid looking, as debris or other obstacles do stop humans in emergency situations.

22

23

I don’t mean “why” at the high level of “why do this at all.” I meant the specific “why” for a particular robot - what do we want this robot to do. It was intended as an answer to the question of what you were missing from the list of things a robot needs.

In addition to the list you provided, a robot would also need a defined purpose, and guardrails to confine them to that purpose.

24

I don’t think it’s always necessary to define the ‘why’ - there are pieces of research and development that create general-purpose solutions that find the problems they are supposed to solve after they are made.

There might be fuzzy ideas about possible applications of a technology, but not every solution begins with a well-defined problem, and some technologies expand into areas beyond their original concept. LLMs, for example, were first devised for language translation.

I don’t think this thread needs to consider the ‘why’ - this is about the ‘what’ and the ‘how’, and perhaps the ‘when’ - besides, given that plenty of different development teams have already worked on humanoid robots, and plenty of writers have conceived of imaginative futures containing artificial people, the ‘why’ can be left as ‘any of those reasons, whatever they are’.

25

I was going to start a thread about this. I don’t think most here realize what leaps have been made with bipedal robots. Boston Dynamics is actually behind the game now. There are five or six other humanoid-type robots under development now, and some already at work in factories.

Adding an LLM to the mix changes the game. Until now, programming a robot meant programming every step it made. AI is in some robots, but usually just as part of an object recognition system or something. I did some work in a factory that had a robot for inspecting the paint job on parts. It was quite smart and had a vision system that could spot all kinds of defects, but you still had to program every movement.

With an LLM behind a robot, you can program it in English, and it can come programmed with a rich set of generalized, pre-trained tasks. So ask one of these robots to make you an omelette, and it will find the kitchen, find the fridge, locate eggs, take them out, find a bowl, crack the eggs into the bowl, etc. It needs no specific programming to do a wide range of tasks.

One of the robot manufacturers has built a simulation for training it. Simulated package pickups, doors to open, etc. The LLM practices until it can do it all flawlessly, then it’s moved to a real robot which can now carry out those tasks in the real world.

Here’s Tesla’s Optimus Robot walking across a factory floor:

Optimus can self-calibrate by just looking at its hands and feet and waving them around. It has sensitive feedback sensors in its ‘fingers’, and can handle everything from heavy weights to cracking eggs into a bowl.

OpenAI isn’t being let out. They just invested a huge amount of money into Figure Robotics to create an LLM powered android robot:

That’s really not the case anymore. Most of the humanoid robots currently have endurance in factory measured in a few hours, with at least one being able to do an 8 hour shift. The Tesla Optimus has a 2.3kWh battery, which Tesla says will allow it to operate for a full working day, depending on the activity. Also, if a robot can recharge in half an hour it’s no different than a human taking a lunch break.

26

Wow. I sure didn’t. Thanks for the updates / corrections.

27

Do the robot minds actually have to live inside the robot bodies (and therefore consume power carried by the body)? The “mind” could be a program that analyzes sensory inputs from the robot while living on a desktop computer that’s plugged into the grid just as easily as it could live in a minituarized computer in the robot’s torso or head.

In fact, living on a desktop would probably be better for networking multiple robot bodies together.

28

The change has been really fast, so it’s not unreasonable to be behind on the news if you haven’t been speciically following it.

The real killer market here is health care. Both home service and in hospitals/clinics. As the population ages, there is going to be a huge amount of pressure on health care. But if a humanoid, intelligent robot can be purchased for $20,000 (Tesla’s goal), a retired person could have a robot that will cook, clean, help them with daily tasks, monitor their condition, order groceries, and be a companion. There’s a lot of lonely old people out there, and the Japanese experience is that they will bond with robots and feel better.

In a hospital, a humanoid robot on wheels can be a reall labor saver for nurses. They are already being implemented. LLMs with supercharge their abilities.

Humanoid robots are currently a 1.8 billion dollar industry. But it’s expected to grow to 13.8 billion by 2028.

29

The robots communicate through Wifi mostly. And I think the current LLM based ones are probably using an API to connect to a commercial LLM.

But in the future it’s almost certain that they will have self-contained brains. The LLM you need to run a robot is likely to be much smaller than something like ChatGPT. Google has a Robot controller called RT-2 designed specifically for the task. One version has a 12B parameter LLM, and the other a 55 Billion parameter LLM. ChatGPT is 180 billion. The 12B LLM should run just fine on a good single board computer with 24GB of RAM. Also, their models are general purpose. If you need a robot for a more specialized task, it might be able to use a much smaller LLM. A 7B LLM will run on a phone.

And remember, the real processing cost is during training. A pre-trained robot would only require inference, which can be done very cheaply in smaller models.

Exciting times. Most people have no idea what’s coming in the next few years.

30

I’m impressed. At least by their claims. Even if exaggerated they may have cleared a hurdle to have humanoid robots perform useful labor. There aren’t that many steps before robots can build more robots. That doesn’t require high levels of autonomy, just enough to avoid detailed programming of every action.

31

I don’t see how it’s a hijack. You were arguing that an LLM could provide volition to the robot you created. Thus discussing whether an LLM can actually provide volition is relevant.

The current models that Sam describes seem to be more about using an LLM to program the bots, not to give them their own volition. Volition would mean the robot was choosing what to do.

That seems far more promising to me. I’ve seen an LLM be programmed to act like it is in an RPG, and it seems that they do start going off the rails after a while. It would be best if that isn’t done in the real world.

Because that’s important too–any robot needs to be controllable. We’re not creating actual people here, creating life. That would be monstrous, since it would then become immoral to end said life.

32

Considering that exaggeration is the norm with Tesla and others, I would expect to see a change in the labor front to take longer. Particularly, because of the price range issues.

I have to say that there is a bit of disconnect here, the OP is dealing with autonomous robots, I don’t think those are gonna be apppreciated by corporation bosses.

33

OP, if autonomous robot people could be constructed right now then why haven’t we seen any?

Those who want to make claims are thick on the ground. All those endless press releases about flying cars, e.g. Those technically exist, but are no more than prototypes that go nowhere except out of business. Or the claims about making human clones that made headlines a few decades ago.

My point is that if these autonomous robot people could be made today, somebody would be out in public loudly claiming that they’ve made one. All of human history since the start of the industrial revolution points to this. Yet, nobody has, to my knowledge or presumably yours since you started this thread.

Ipso facto, autonomous robot people cannot be constructed right now.

34

Autonomous Robot People? Ah, Postal Workers.

35

There is often a large gap between when the basic technological pieces have all arrived vs. when someone actually decides to build the thing.

Useful bipedal robots require several hi-tech components: power-dense motors, energy-dense batteries, extremely fast computers, high quality digital cameras, the software stack (particularly LLMs and other machine learning systems), and so on.

These things have only come together very recently–especially for a reasonable price. And so the various projects have also only started recently. If someone could wave a wand and magically skip the rest of development, we could have the robots today. That’s not possible, but it doesn’t mean the OP’s premise is wrong.

Another video of the Optimus robot:

That was an older version of the robot, but IMO the moves it showed are more compelling. It clearly has a degree of proprioception (awareness of the body’s position, movement, etc.).

Figure is doing interesting things, too. They showed their robot how to operate a coffee maker… and it learned how to operate the coffemaker:

36

It’s going to be difficult to drive the price down, but there is nothing fundamentally expensive about the robots. They’re a hunk of mostly aluminum and other cheap metals, and computing/sensor devices not too different from what cell phones use.

Tesla at least is shooting for $20k robots. That’s several years off at least, but if they can get anywhere close to it (and the robots are useful), there’ll be an enormous market for them. Even upper-middle class households will be able to afford them. People buy $1000 Roombas today. But a bipedal robot can do 100x the work of a robovac.

37

I think end-to-end training may prove even more valuable. Language support is crucial, of course. But you also need to be able to show the robot how to do something, whether it’s where to put your dishes or how to scoop the litterbox. Having the robot be able to copy an action just based on visual input will make it far more valuable than one simply trained on canned routines (that said, a robust collection of canned routines is also valuable: imagine a robot that already knows how to make any of a million recipes, because the training was amortized across the whole fleet).

38

We’re going to need some way to regulate the behavior of these autonomous robot people. I suggest a short list of generally worded principles that sound comprehensive at first but are in fact open to broad interpretation as evidenced in a series of logic puzzles disguised as short stories.

39

It’s possible that I was hasty in dismissing that avenue of discussion, but in every thread I have seen where it is discussed, it just devolves into irrelevant (and often misinformed) bickering about whether anything other than a human can have sentience, whether algorithms for AI/ML can do anything more than ‘they were programmed to do’, whether LLMs are ‘just’ elaborate autocomplete software.

I don’t think it matters, because considering the state of technology as a phenomenon, there are already systems available that can function as though they have the capacities of reasoning and intelligent decision. For this thread, it’s irrelevant how they do it, and inappropriate to argue that they can’t do it, when it’s already been demonstrated in practical examples.

40

Not necessarily. Running everything over a wireless connection does introduce some latency, but probably not enough to worry about, however a robot that is entirely operated by a disembodied controller would just stop working if the connection is lost. It may be that some sort of hybrid setup could be implemented where there is sufficient local function to perform tasks that had already been decided, to maintain housekeeping, avoid obstables, react to changes, etc, with the non-local machine performing the higher level decision making, more complex perception, etc.