Turns Out, Robots Need to Dream Too

[00:00:00]Humanoid home assistant robots are set to become commonplace and it turns out they are going to need to both sleep and dream in order to function. I find this area fascinating because we kind of just lobbed a bunch of language at a robot and to our great surprise, it worked.

[00:00:13]However, we're finding that we need to make it more like us in order to have it function well like us. Finding out that dreaming improves performance and that it may even be necessary for performance is something that we found several times over. Not just in models that remain in your computer, but it's even more necessary when it comes to embodied models. It is believed that in humans, dreaming serves a couple of purposes. It allows It is believed that in us, sleeping serves a couple of purposes. It gives our brain a break and we can clear out toxins that build up over the day, but it also helps us consolidate memory.

[00:00:46]At least that is the theory. In robots, it also serves several purposes. One of which is that it prevents catastrophic forgetting in artificial neural networks. So if they just operate continuously, something will happen and they just kind of forget what they're doing. They forget context. But if they're allowed to sleep and have similar functions as dreaming, they don't. Sleep and dreaming gives them more than just a space to consolidate the memories that they have, although that is helpful so they can figure out what went wrong and what they can do better in the future, but also a space to imagine future events that they've never had before. This is also a great way to do software updates. So say a company wants to do an update where a robot can now change your oil. They can then send that over and the robot can dream about it so it will wake up knowing how to do it. It's actually fairly effective. But the larger point is that robots need world models. We learn over the course of our lives what works and doesn't work and then we change your behavior according to it.

[00:01:38]And when you reach for a glass, you're also calling on your knowledge of physics from your lifetime. You know how to reach for that cup. Robots don't have that. And many of the models that we have are simply reacting. So they have a sensation and then react to it rather than figuring out what a likely sensation will be. We hallucinate the world around us. This is why eye puzzles work. You have what you think you're going to see and then you have what you actually see. And those two things do not always align. This research group is calling what we already understand as intuition and world modeling as touchdreaming. Incidentally, this is also considered to be one of the broad signs of consciousness. And if you've listened to me, you know that I find that the definition of consciousness is a little bit wishy-washy. So, you're going to have to model what you are going to experience. You also have to model what is going on in somebody else's head, which AI actually does pretty effectively. They have to figure out what you want so that they can react to that. And some of that is applied to guardrails, and that does bother me.

[00:02:38]This also requires broad sensory perception. So, a robot is not just going to have to figure out what will likely happen. It's also going to have to figure out where it is in space, what it's doing. So, they also suggested having additional cameras on the hands.

[00:02:51]One thing that I really like about this area of science is moving from teaching robots how to behave from a human being doing it. Because a human's body is not going to look just like a robot's body, and it may not be applicable to lots of different robots. They have to learn through trial and error in their own bodies in order to move effectively. And yes, that is incidentally how humans develop. Robots are mirrors. Okay, we only really have ourselves to reflect upon in order to create something and we assume that we are the most efficient way to operate. So, we create systems that operate very much like us. Now, robots are still pretty expensive and also kind of dangerous. Some of the cheapest ones that I've seen being marketed as assistants run $8,000, but that's on a very cheap end. Most of them are going to run 20 to $30,000 or more.

[00:03:36]So, having one in your house is not really something that the average person wants to do yet. If they do essentially become like an additional person, then yes, I could see them being widely available and widely purchased. I would really not think that robots would catch on all that quickly, except for how quickly AI caught on. I mean, it's basically in every aspect of our lives.

[00:03:57]And I think people are hoping that robots will follow suit, that they can be extremely useful to do all the things that you do not want to do. So, will they eventually become like having a cell phone? I think that is a possibility, but I also think that companies are overpromising. The technology really has not kept up as quickly with the AI side of it. So, we still see robots trip and we really need to improve the embodied forms before we start putting them everywhere. Although, I'm also concerned about the things that people will do to their robots because I understand people. Either way, I really like the concept that AI needs to dream in order to function and I will keep you up to date. I hope you've enjoyed this.

[00:04:38]Follow for