Training robots to have human-level intelligence
At the Samsung CEO Summit in October, Vicarious CEO Scott Phoenix discussed his company’s plan to bring a new, more advanced level of artificial intelligence to robots. By doing so, he believes we can make robots less expensive, more affordable, and eventually ubiquitous in our lives.
To understand Vicarious’ motivation, you first have to understand what Phoenix sees as a paradox in the fact that the world is full of really cheap robot parts, yet nobody owns any robots. Meanwhile, those robots which do exist today are only capable of performing the singular tasks that they are programmed to accomplish.
“We thought that we would have general purpose robots that could do all kinds of different stuff, without being hand-programmed to do each task, one at a time,” Phoenix said.
Phoenix compared the state of AI today to the way that intelligence has evolved in animals over the last 600 million years. If we go back that far, we encounter the first so-called “intelligent animals” whose behaviors are based on instinct developed over centuries of stimulus and response.
Most animals exist in an “old brain” paradigm that enables them to perform relatively advanced tasks like hunting for food and navigating their environments. But those animals are not very adaptable to new environments and have difficulty learning new tasks.
Furthermore, they need a lot of training data — a couple hundred million years in some cases — and can be easily fooled. According to Phoenix, an “old brain” animal “isn’t going to have a model of the world of itself, of what it’s doing. It’s going to give the illusion of intelligence without actually being smart in the way you and I think about being smart.
“It’s remarkably similar to what you see in today’s AIs, where you have AI systems that give the appearance of intelligence,” he said.
Artificial intelligence can be trained to play Atari games, chess and Go, but it usually takes a lot of training data. Those AIs don’t generalize well to new tasks or changing conditions, and can be easily fooled. That’s even more true when AI is applied to robots, which have to operate in three-dimensional space.
“To get to that world where we have billions of robots, we really need to switch models to something that can learn causality and can understand how to control itself in a changing environment,” Phoenix said. “That means to shifting to something that works like your brain and mine.”
By building a model of artificial intelligence that operates more like humans and other mammals, the hope is that Vicarious could provide a new level of adaptation, planning, and generalization that requires less training data and unlocks new capabilities.
“By adding a layer of intelligence… you don’t need such expensive parts and so you can drive robots from being something that are very bespoke and very expensive, into something that’s much more in the scale of consumer electronics,” Phoenix said.
With more intelligence, robots become capable of doing more, which has the economic chain reaction of making them more desirable. That will eventually make them less expensive and more affordable, with increased volume driving down their cost over time.
That will lead to a future where robots “are inexpensive things you see everywhere and they have general purpose hardware and they run intelligent software that lets them accomplish whatever tasks that they’re physically capable of,” he said.