Autonomous robot people could be constructed right now

In My Humble Opinion
1

I think all of the parts necessary to build autonomous robotic intelligent people already exist, and although integrating the parts may present some technical challenges, the hardest parts have already been invented.
I suggest that an autonomous being needs the following features in order to function:

  1. Senses to receive inputs from the outside world
  2. Perception - the ability to model the environment and understand the raw sense data
  3. A body physically capable of locomotion, object manipulation and proprioception
  4. Motor skills to be able to move the body in an organised fashion and drive the manipulators
  5. Intent, agency and volition - the ability to decide to do things

Senses are very significantly a solved problem; electronic cameras and sensors to measure pressure, temperature, battery level, etc already exist in wide diversity.

Perception is solved - we have image recognition algorithms, including neural models that can be shown an image and will accurately and reliably describe what it sees in detail.

Physical body - see Boston Dynamics.

Motor skills, I think are solved - there are vision-assisted algorithms that will take natural-language instructions (such as ‘place the green box on the blue tray’) and translate those instructions into coherent movement of manipulating motors.

Which leaves the thinky part. Intent, agency, volition, etc - but this is actually solved using LLMs - people often say that LLMs don’t have any agency, but it seems like it would be incredibly simple to just make that happen. In fact you can do it already yourself - you can (for example) tell ChatGPT ‘For the purposes of this conversation, I want you to roleplay the intelligence of an autonomous robot; I will describe your sense and inputs; you just need to tell me your desired actions’ - and it will act as if it is the player in an RPG.
You do need to keep asking it ‘what do you want to do next?’, but that could literally be performed by a do-while loop that keeps feeding it the descriptive version of the sense inputs (that is, the textual output of the ‘perception’ model), with the text ‘what do you want to do next?’ appended to it - then the output of this LLM can be fed to the ‘motor skills’ model and acted upon.

I tried this a little bit - basically just me as the DM and ChatGPT as the player, in an imaginary domestic setting, starting in the kitchen of a house; first off it seemed to think (without me saying so) that its purpose was that of a servant, so it cleaned and tidied; this wasn’t going anywhere, so I let it discover a note telling it that it was the owner of the house and could do anything it chose to do; it went outside, looked around the garden, studied the plants and wildlife, rescued a drowning hedgehog from the pond, then wanted to continue exploring, observing and doing stuff.

I’m not sure what would happen in an extended session like this - LLMs have a tendency to go a bit insane after a while, so maybe it would eventually go on a killing spree or something.

Of course, I am not arguing that such a machine would have any inner experience of consciousness or anything like that, only that the parts necessary to build it and let it loose all pretty much exist already.

What have I missed out?

2

The “why?”

Are we doing it just to do it? Are we creating a new slave race? What’s the point?

3

I don’t think that’s really in the scope of the thread. Why do anything?

4

You can’t make “Your Plastic Pal Who’s Fun to Be With” without a “Genuine People Personality”.

5

That’s also possible; you just have to choose between Microsoft racist or Google racist personalities.

ETA: better add some context…

(Google’s Gemini model recently ran into problems where it was trying so hard to be unbiased, it accidentally a racism)

6

We’re not there yet. Not really much closer than we have been in the past. In recent years there have been developments in smaller electric motors, batteries, and sensor devices. But the best we could do is make an incredibly expensive device that breaks down often and doesn’t perform much useful work, none of which can be performed more effectively or comparably to the cost of human labor.

We will get there. Younger folks should enjoy life now.

7

We want that? Really? For sure?

In my world, the purpose of robots would be to more than competently perform important assigned tasks. I don’t want them to be autonomous in the sense that they can dream up things to do and then decide to do them. We don’t need more people, not even synthetic ones.

8

The thing that gets me about that is seeing the right-wing sources of info being insulted, I mean, of course it was bad to have black nazis, but the use of the already old dog whistle (woke) points at them acknowledging that they like their Nazis to remain race pure.

9

According to the tag on this thread, evidently you’re missing fashionable attire for the robots.

But really, I think the answer is actually “why”. An empty-slate robot has no defined purpose (and therefore no reason to be built), and at best would be unpredicatable, and at worst a horrible and potentially murderous entity. So anything we build needs to have a defined purpose with all the necessary guardrails in place.

10

Pretending you’re serious …

What’s missing is power supply. A humanoid robot (“mentally” autonomous or not) built to the lowest possible power consumption given current tech could not begin to be strong enough to drag its own batteries anywhere.

11

Yes they can. Boston Dynamics ‘spot’ exists, for example.

Atlas is more humanoid, but can only get an hour of runtime between charges, however there is no specific requirement for these things to be completely humanoid.

12

Large Language Models have zero native intelligence. They’re robot parrots that can regurgitate a trillion canned responses, which is not even close to intelligence.

13

This is irrelevant, as well as not particularly correct.

14

It becomes relevant when the thing that up until now has been doing a fair job of tricking people into believing it’s actually sophont suddenly does something so appallingly mindless that stupidity is an insufficient term.
I’m not saying that true AGI is impossible but LLMs aren’t it.

15

Please desist the hijack. There must be plenty of other threads where the question of what really is intelligence and does ai have it, is on topic.

16

I think there might be an issue still with the interaction between sensors and motor skills. I may be completely off-base but I think there’s a degree of complexity involved in interpreting a constantly updating 3D image and understanding that the green box continues to be the same green box even when a cloud passes over the sun etc. that we might not quite be at yet.

I’d also be interested to see what happens when the Autonomous Robot Person starts integrating different types of sensory input into a coherent whole - the smooth, heavy, green box, for example.

But I suspect we might get over these hurdles (if they are hurdles) soon enough. It’s the volition part that I think will be stickier. First of all, LLM volition is a very limited kind of volition. The ARP has to be told what kind of volitions it has (servant vs houseowner) and it has to maintain these consistently. I think that consistency is the hard thing. LLMs can take as input the entire conversation that you’re having with them, but I’ve seen various examples where even so it cannot “hold on to” an idea consistently and appears to forget things it knew a second ago. I also believe there is some sort of limit on how long such a conversation can get, which raises the possibility that the ARP will be like Leonard in Memento - every n minutes it forgets what it was telling itself n+1 minutes ago and effectively only has a short term memory to operate on. Which would lead to erratic behaviours and the inability to carry out plans.

None of which is to say it wouldn’t be cool to try!

17

I will go for what I think it will be the main reason: understanding human physiology and how the brain interacts with it. I remember how in the past the need to understand how and why new products could be harmful to humans got some people to hire volunteers as guinea pigs early in the past century, (can’t find it right now, but the British QI TV show reported about a “belly brigade” group of volunteers that consumed dangerous products to see how humans could be affected) of course there have been very unethical experiments done to humans, particularly in times of war.

AI proponents like Jeff Hawkins pointed a while ago that we will not get robots like C3PO, but there will be AI, reason being that humanoid like robots don’t really make sense economically or in practice. Unless… (IMHO) medical groups with deep pockets realize the advantages of combining human components with the robot beings so as to test for simulated conditions or cures. And to check for what could get humans sick with fewer ethical issues related to human experimentation.

(Yes, that will go for sex too, but in this context, it will be to check for venereal diseases rates, infection vectors and cures. less issues with ethical considerations here.)

18

I’ve seen technical demos where that does seem to have been solved, but the perceive-decide-act loop might need to be tighter and more finely grained than what I described in the OP (a model describing the world to a model deciding what to do, which in turn instructs another model to do something practical)

19

One quite compelling way that humanoid robots do make sense is that they would fit well into the existing infrastructure, which is currently configured to fit humans to a significant degree. A robot wouldn’t need specialised tools or accessories if it could just pick up human tools and use them, walk through human-optimised doors and climb up human-size stairs, etc.

20

I don’t think the fact that some people ask ‘why?’ is going to deter the people who are driven by ‘why not?’