Humanoid robots are the next step for AI. Here's how to train yours

The next phase of artificial intelligence (AI) is robots, which will help with the global labour shortage, an Nvidia executive told Euronews Next.

"We are at a very interesting point in time. The promise of robotics has existed for a long time. It's been in our imaginations and science fiction," Rev Lebaredian, vice president of Omniverse and simulation technology at Nvidia, told Euronews Next at the Computex technology fair in Taiwan.  

He said that despite tech companies trying to build a general-purpose robot for years, the issue has been that, despite being able to build the physical robot, programming it has always been a challenge.

"AI has changed all that. We now have the technology to make robots really programmable in a general-purpose way and make it so that normal people can programme them, not just specific robot programming engineers," he said.

Companies such as Tesla are racing to build humanoid robots and have made strides. Last week, Elon Musk’s company said its Optimus robot had learned to perform household chores.

However, there is still much for robots to learn.

For Nvidia, the company says robots should learn their tasks in the virtual world for safety, but also because it would take too long to train robots with humans. 

"The only way to actually create these robots, intelligent ones, is to employ simulation," Lebaredian said.

"The fundamental problem that we have with physical AI is that AI is data hungry. You have to feed into your AI factory lots and lots of quality data to give it life experience to train from".

He said that with large language models (LLMs), there is a large amount of data online to train them.

Training your robot with data

But he said in physical AI, there is no such data that can be mined.

"To get all of the information we need to train a robot on how to pick up an object, we have to go create it somehow," he said. 

"Collecting it from the real world is not possible. We can't create enough data. Even if you can, in some cases, it's dangerous, it's time-consuming, and it is expensive".

What is needed is "a way to go from fossil data to renewable data sources," Lebaredian said. And the best renewable data source for physical data is a physical simulator, he added.

Once your robot is tested, or has "graduated" and looks like it is working well, it can then go to its first employer. 

"A new college graduate is trained on a corpus of publicly available data. You study from textbooks and information that everybody has access to everywhere. And you have a generalist that enters your company, and they're useful," Lebaredian told Euronews Next.

"But they're not really useful until you train them for a few years on the specific proprietary information and data in your company that's about your domain and your particular practices and how things are done," he added.

In robot terms, it means that you could then specialise your robot with your own data to make it work best for you.

Lebaredian did not specify a date when humanoid robots would come into our lives, but he said it would be "soon".

Where and what to use your robot for

The first use cases for them would be in factories and warehouses.

"I think industrial use is going to be the first one because even if we can build a perfect robot that you can use in your home, it's not clear that all humans will want one," according to Lebaredian.

"But industry, there is a great need for it. There aren't enough young people replacing the older skilled workers who are retiring in every country".

Global labour shortages have reached historically high levels in the past decade, according to the OECD.

Population declines, as well as ageing populations, and the fact that many people do not want the "three D" jobs, which, according to the Nvidia executive, were "jobs that are dangerous, dull, and dirty".

Taiwan has jumped on this robotics need and is set to launch a five-year plan to boost the robotics industry in a bid to plug labour shortages, the government announced last week.

Taiwan’s population decline would strain the economy and the nation’s ability to care for vulnerable and elderly people, Peter Hong, who heads the National Science and Technology Council’s (NSTC) Department of Engineering and Technologies, was reported as saying, according to local media. 

Lebaredian said that after factory use, humanoid robots could help in retail, as he hears a lot of companies saying they cannot hire enough people to stack shelves.

He also said they could be used in mines, nuclear reactors, or even in space. Eventually, he said they could be used to take care of the elderly if the demand is there. 

How to make your robot safe

But just as we get excited about this next phase of AI, LLMs are still getting much wrong, which is causing them to sometimes make things up. Errors caused by a robot in the physical world could be much more dangerous. 

However, Lebaredian believes that just like autonomous vehicles seem like science fiction at first, people eventually get used to them, and the technology improves.

"In generative AI, yes, there's still some stuff that's inaccurate, but I think you have to admit, in the last two and a half years since ChatGPT was introduced, accuracy and the quality of what it's producing have increased exponentially as well," he said.

But he added that perhaps chatbots will never be quite right because we want humans to perform the tasks. 

"There's actually no right answer for a lot of that stuff," he said.

"But for tasks that we have in industry, that is actually something that's very measurable, for example, did it accurately pick up this object and move it over here and do that safely and robustly?"

He said those systems can be created, tested, and made sure they are safe before deployment.  We can create these systems, test them, and make sure that they're working well before deploying them.

"We have machinery and systems that we create that are quite dangerous if they're not set up right. But we've managed to create nuclear reactors and these systems, and keep them safe somehow. We can do the same with physical AI," he said.