The Global Project to Make a General Robotic Brain

I wasn't aware of this, but from wikip:

"In 1973, professor Sir James Lighthill was asked by the UK Parliament to evaluate the state of AI research in the United Kingdom. His report, now called the Lighthill report, criticized the utter failure of AI to achieve its "grandiose objectives". He concluded that nothing being done in AI could not be done in other sciences. He specifically mentioned the problem of "combinatorial explosion" or "intractability", which implied that many of AI's most successful algorithms would grind to a halt on real world problems and were only suitable for solving "toy" versions.

The report was contested in a debate broadcast in the BBC "Controversy" series in 1973. The debate "The general purpose robot is a mirage" from the Royal Institution was Lighthill versus the team of Donald Michie, John McCarthy and Richard Gregory. McCarthy later wrote that "the combinatorial explosion problem has been recognized in AI from the beginning".

The report led to the complete dismantling of AI research in the UK. AI research continued in only a few universities (Edinburgh, Essex and Sussex). Research would not revive on a large scale until 1983, when Alvey (a research project of the British Government) began to fund AI again from a war chest of £350 million in response to the Japanese Fifth Generation Project (see below). Alvey had a number of UK-only requirements which did not sit well internationally, especially with US partners, and lost Phase 2 funding."

https://en.wikipedia.org/wiki/AI_winter

Seems this is the report if anyone is curious to read it: https://www.aiai.ed.ac.uk/events/lighthill1973/lighthill.pdf

Pretty much the only "AI" in that demo is edge detection (against a uniform background) and pattern recognition (by direct comparison with a handful of simple geometric shapes). Other than that, the robot seems to be following a hand-coded set of instructions to assemble the provided parts.

The current state of research, as described in the article, is a neural network which recognizes which hardware it's currently controlling and uses it to autonomously perform tasks assigned using natural language. So it seems to me that there has been some progress.

As far as I can tell, all the AI stuff happening at the moment seems to be matrix multiplication. Someone wrote up the equations behind transformers the other day and it was a half-page of maths.

I skimmed the video and I think the point you're making is not much has changed? The important variable is FLOPS/$, which is nothing like 1974 and still increasing exponentially. The researchers have known what to do for decades; it is only a question of being able to actually do it.

So most the specialist problems we've had that were not 'generalist hard' we've already mostly solved with machines that we don't consider complex robots.

It shape of the robot isn't the issue, the problem space of reality is. A roomba can suck stuff off a floor, it has a much harder time dealing with (or even identifying) a sock that needs moved out of the way first. To do that for all the different objects that could show up in front of you, you need a general purpose AI.

> While this is pretty cool I believe generalist robots will not have a fraction of the skills an average human has , at least not within a realistically foreseeable timeframe, let’s say within this century.

I'm pretty sure they will be flipping burgers by the 2040s, if not before, and doing everything else needed to efficiently make a (big) percentage of people jobless in first-world economies. Though, not in parts of the world where electricity regularly in the wire is still rare...because those problems don't solve themselves in a few short decades.

Now, we both are cynics and should go for a beer together. Who knows, maybe we will come up with a more catastrophic and highly probable scenario that combines your outlook with mine...

The problem is that even a specialist vacuuming robot needs to be more of a generalist than you realize.

For example, take the "nothing was allowed to be on the floor" restriction. To relax this restriction, the robot needs to know what it can and cannot do for something it sees in the floor. List everything that could be on the floor (I'll wait). The robot needs to recognize all of these things, and know the correct behavior for each.

You could still do this! You'd need to label a ton of items and hardcode a bunch of behaviors. After all this R&D your robot vacuum would need to cost $5k-$10k, and you'd wish you'd worked on a higher priced product like a robotic forklift instead. Still, it's feasible to build this.

However, manipulation is a few orders of magnitude more complex than navigation. You have to recognize many objects, their precise poses, and many aspects of the objects. Think about opening a can with a can opener. The robot needs to recognize a few parts on the can opener, and how it fits on the can. Then you've got to hardcode behaviors for attaching the opener and then turning the knob until done, and removing the lid. Doable, but very very hard.

This is feasible, and you can build a can opening robot, but after 9 months of R&D, that's all your specialized robot will be able to do, and oops, there's 40 more tasks it needs to accomplish to cook a dinner. The only way to build this product is to tackle all the tasks at once, and that's why this research is so important. Everything you want a robot to do needs O(dozens) of individual tasks, and when each task takes O(year) to build it's impossible to finish.

Good floor cleaning robots are effective. Your anecdote doesn't prove otherwise. It's like a 4-5+ billion dollar market.

If you look at demos from this article, ALOHA, Tesla's bot, Asimo, etc. you will see more and more skills being demonstrated.

The biggest thing missing from most systems is probably strong fluent motion with integrated sensing and actual hands. But the Tesla robot for example has made a lot of progress on those fronts and does have hands with touch sensors.

What are the skills that your average human has? And let's see if we have robots that can do that. Maybe not quite as well YET, but still the same skill.

- can walk around. Check

- play soccer. Check

- locate and pick up objects from a flat surface. Check.

- assemble objects together. Check.

- put dishes in a dishwasher.check

- play the piano.check

- climb stairs. Check.

- do a front flip. Check.

- open doors. Check.

- drive a car. Check.

- write a computer program. Check.

- draw or paint. Check.

- put clothes in dryer. Check.

- understand and produce natural language. Check.

We are really at the point of just making these things work better and be deployed. And there is rapid progress.

The large multimodal models provide a new level of generality that is accelerating this.

I believe that within say 2-5 years you will see an explosion of robots with more skills than any human could ever hope to achieve in their lifetime. Just like you can download LoRAs that make LLMs more capable in a certain programming language or art style, there will eventually be adapters to instantly provide any type of skill desired.

You will just say "I would like a martial arts lesson" and your android helper just looks off into the distance for a few seconds and then says "I.. I know Kung Fu!" Then launches into a demonstration worthy of The Matrix.

My layman understanding is 1. big, practical size robotic arms are expensive and 2. also very slow, making real training expensive and time consuming. And 3, simulations aren't high-fidelity enough to be fully sufficient for training, but they definitely can help.