Ask HN: When do you expect ChatGPT moment in robotics? | Hacker News Reader

Ask HN: When do you expect ChatGPT moment in robotics?

10 pointsp1esk20 comments

Current humanoid robotic assistants are in early stage - somewhere around GPT2 level - they're starting to perform very simple, very narrow tasks, but stumble a lot, and still cannot do much. However, I've been tracking the progress in the last couple of years, and I feel that GPT3 level might already be happening, and some startups demonstrate impressive things (e.g. look up Generalist AI or Physical Intelligence). Plus the funding all these startups are getting should allow them to scale their methods 10x-100x of what has been tried so far. I'm not sure any additional research breakthroughs are actually needed to make the leap to usable products.

Therefore, we might soon see a ChatGPT moment in robotics - a commercial availability a physical robot that will be capable of performing useful tasks: cooking, cleaning, simple repairs, yard work, elderly care, etc. Just like ChatGPT-3.5, this won't be reliable, and the robots will still be clunky/dumb, but I think it will be obvious there's a step change/phase transition, where most people realize a paradigm shift is happening. Soon after that initial stage, it will lead to something globally transformative (like GPT4): think of how software engineers currently using Claude Code, but applied to physical world, for everyone, everywhere. Well, everyone who can afford a robot like that - I'm guessing it will cost like a premium car.

I'm curious when this will happen, and what will be the short and medium term consequences of having physical world assistants? My intuition is there's 40% chance we will see it this year, and 70% by the end of next year. I'm pretty sure (90%) we will have somewhat useful robots in people's houses within 3 years. I do realize this might sound very optimistic, but it would had been just as optimistic to predict ChatGPT two years before it was released.

"the funding all these startups are getting should allow them to scale their methods 10x-100x.." .. "Therefore, we might soon see a ChatGPT moment in robotics" -- I don't think so and no, the second statement is NOT entailed by the first. Why would it? Because 100 is a big number? Do you have any idea how much more data LLM needed to be trained for a GPT3 level compared to the data available for robot training right now, and how low dimensional the space is in which LLMs operate compared to robots?

"My intuition is there's 40% chance we will see it this year" -- again, why? Don't you realize that people have been working in robotics for 65 years, and these people don't live under a rock either. They knew about GPT3 because 2023. So why is it NOW less then 10 month you think that this breakthrough will happen?

Can you share more about data availability part?

What kind of data do you need which is missing now, isn't simulations enough? curious to learn more about bottlenecks in general

I think it’s a combination of simulation, YouTube videos, and specially recorded training footage. The last one is expensive, but given the funding these startups receive, I’m pretty sure they can scale their RL methods at least 10x.

One big difference between LLM progress and robotics is that language models benefited from purely digital feedback loops — training, testing, and scaling could all happen in simulation.

Robotics still pays a heavy “reality tax”. Every improvement eventually has to survive messy, unstructured physical environments where sensing, actuation, and safety interact in unpredictable ways.

My guess is we’ll see a ChatGPT-like moment first in semi-structured environments (warehouses, logistics, industrial assistance) rather than homes.

A general household robot feels closer to a GPT-4 equivalent problem than a GPT-3.5 one because reliability matters much more when failure affects the physical world.

I think there will be a sort of ~meeting in the middle. I.e. when friend was furnishing his appt. he chose specifically the sort of furniture that will make it easier for his robotic vacuum to work. I wouldn't be surprised if there will be some extra allowance for autonomous vehicles comming in next ~5 years. Stuff like that.

Arguably robotics is already happening. Most factories today rely heavily on automation and robotics. We beginning to see the first production-ready self driving cars. We even have some fairly decent humanoid humans now.

The main issue is that the world is complex and it's hard to build a single robot with enough flexibility that can preform a broad set of tasks.

If the question is specifically when can I get a robot that's going to do almost everything I can do around the home, then the answer is probably not for a decade or two. But over the next decade we'll increasingly see robots being deployed in the real-world to solve specific tasks. For example, I strongly suspect postal delivery will be fully robotised by the end of the decade.

Not anytime soon. The big leaps in LLMs don't really carry over to robotics much, aside from some computer vision stuff. I'd say we're still 10+ years out from robots cooking for you. Maybe we'll get some kind of dedicated 'cooking machines' if there's money in it, but not humanoid cooks walking around your kitchen. And even that's being pretty optimistic because right now there's no clear way to get there, and we'd need some real fundamental breakthroughs to make it happen.

we'd need some real fundamental breakthroughs

What sort of breakthroughs are you thinking of?

No idea, and I don't think anyone does, hence the prediction that it won't happen anytime soon. Either we need perfect simulations (seems almost impossible) for training at scale or fundamental algorithmic breakthroughs in learning from very sparse data that we can collect in the real world.

Why can’t it learn from, say, YouTube videos?

sounds unrealistic to me. For military and industrial use cases yes but I think we are at least 5 probably 10 and more years away from that moment. I think there is not enough data, and hardware is not software so the scaling will be more difficult and use cases are so wide that I can not imagine a humanoid cooking and mowing the lawn and repairing the roof. Look how long the autonomous driving takes and this is just moving a car.

will robotics have as large of an impact on our society as software? Almost every industry needs the internet and benefit from software. I actually have a hard time thinking about horizontal use cases for robotics in B2B especially.

For example, I’d like to have a farm. But I don’t have knowledge, skills, or time to maintain it. With a robot I can, right?

The cost would have to be real low for a real ChatGPT moment

How low? Would you consider $50k to be low enough for wide adoption?

No, but for industrial and business uses yes. You'd start to see a lot more robots used for warehouses, deliveries, restocking, cleaning, preparing food, etc.

Honestly I’m not really interested in industrial robots. What price would be reasonable for a humanoid robot that can do common household tasks?

I'd say closer to $5000, which is about the cost of a domestic helper for a year or so.

Then again, they don't need sleep, can be jailbroken, they only need closet space, and won't take your money and run. They can do dangerous tasks like fixing the roof too. Menial labor like walking 3 km over to the store to buy me a can of Red Bull. You can have them do pranks at night. They'd probably be valued quite a bit more than a human.

$5000 will still be absolute niche. Cost needs to be close to a phone to get mass adoption. And that might very well never happen.

A robot like that would be far more useful to me than my car (I rarely use a car), and I paid 50k for my car. So for me personally, 50k would be a no-brainer. But of course only if it can do the tasks I mentioned well enough.

Reply on news.ycombinator.com