OpenAI and Sequoia invested 500 million yuan! Geniuses gathered to build a brain for robots!

Bringing AI to the physical world through universal models

Recently, a company that officially announced its existence through Twitter and didn’t even have an official website, Physical Intelligence (Pi for short), made a “glorious debut” with $70 million in financing. Even heavyweight investment institutions such as OpenAI and Sequoia Capital have placed bets on it.

Nowadays, it is as difficult as hell to stand out in the crowded AI field, not to mention getting the approval and investment from the “top AI figures” and venture capital giants.

What is the unique charm of the aggressive Pi that attracts industry leaders to offer olive branches?

The answer may lie in its entrepreneurial philosophy: to introduce general artificial intelligence into the physical world, to build brains and provide power for robots or any physical devices.

A super team of talented people

If you open the official website of Physical Intelligence, you will find that there is nothing there except an announcement. It is as simple as a bare house.

The official account has only posted 4 messages, which seem ordinary and focus on mysticism. But when I click on the follow list, I find that it is reckless.

Professors from Stanford University, University of California, Berkeley, DeepMind researchers, Google Brain scientists... it was a galaxy of shining stars. What I thought was a "star from the sky" turned out to be a high-end game put together by the AI ​​giants.

The company's co-founder and CEO is Karol Hausman, a former senior scientist at DeepMind and an adjunct professor at Stanford University. He mainly studies how to help robots acquire general skills in the real world. He was also involved in the RT-2 and RT-X models, which have caused widespread heated discussion and are known as the "ChatGPT" in the robotics world.

The company's other co-founder, Sergey Levine, holds a Ph.D. in computer science from Stanford and is a "celebrity professor" at the Department of Electrical Engineering and Computer Science at the University of California, Berkeley. He has also worked at Google and is a "top journal maniac" in the field of deep reinforcement learning, with his research cited more than 127,000 times.

Fighting side by side are Brian Ichter, who has worked in Google Brain and DeepMind teams; Chelsea Finn, an assistant professor at Stanford and formerly worked at Google Brain; Wang Quan, a former researcher at DeepMind; Suraj Nair, a former scientist at the Toyota Research Institute’s ML team; and well-known investor Lachy Groom, who will also provide unique insights and valuable experience for Pi’s future commercialization.

"Creator" creates brain for robots

General artificial intelligence has always been the ultimate goal that academia has been longing for. Unlike narrow artificial intelligence, which can only play a role in specific fields, general artificial intelligence can handle and learn a wide range of tasks, even tasks that do not currently exist.

Pi's grand vision is to build a "brain" for robots and give them general intelligence. Just as large language models have brought revolutionary progress to natural language processing, Pi hopes to develop a general artificial intelligence model that can not only understand language like the human brain, but also control robots to complete various tasks in the real world.

However, achieving this ambition is not easy.

Although artificial intelligence has made great progress in areas such as natural language processing in recent years, the development of robotics is still relatively lagging behind. The main reason is that unlike language models that can be trained with the help of massive amounts of text, robots need to collect training data from the physical world, which is undoubtedly a more time-consuming and labor-intensive process.

In the past, although some companies have tried to develop universal robot software, they have failed to make fundamental breakthroughs. For example, Willow Garage was founded in 2006 to build a software system that can be shared among multiple robot platforms, but it eventually closed its business in 2014.

Therefore, Pi's first priority is to break through this bottleneck, and plans to start from the two key areas of scale and generalization.

The first is to collect a larger, more diverse and richer robotics data set. As the company's co-founder Karol Hausman said: "Our goal is to bring artificial intelligence into the physical world through a universal model that can power any robot or any physical device, basically for any application." With the support of large-scale data, AI models will be able to learn more knowledge and skills from it.

The second is generalization. Traditional robotic systems are often tailored for specific scenarios and tasks and lack generality. Pi hopes to develop a basic model that can be applied to various robotic platforms, various physical environments, and tasks. This requires innovative model architectures and training algorithms that combine robot control and large multimodal model training.

Amazing financing lineup

According to people familiar with the matter, Pi has currently raised $70 million from Thrive Capital, OpenAI, Sequoia Capital, Greenoaks Capital Partners, Lux Capital and Khosla Ventures.

Among them, OpenAI and Sequoia Capital are the most eye-catching. OpenAI is the hottest "hot chicken" in the current field of artificial intelligence. It is obviously not limited to the original intention of "non-profit organization". It is constantly expanding its territory and building its own business empire. As a synonym for the current AI technology, its investment targets also represent the most cutting-edge technology direction.

In addition, Thrive Capital, one of OpenAI’s largest shareholders, and Vinod Khosla are also among the investors. OpenAI co-founder Sam Altman and Khosla have close ties, which may also provide some clues to the investment lineup behind PI.

Sequoia is one of the most well-known top venture capital firms in Silicon Valley. Although the saying "Whoever gets Sequoia gets the world" is too exaggerated, the fact that Sequoia is optimistic about Pi is enough to prove that Pi is capable.

However, the competition faced by Pi cannot be ignored. Figure, which is also committed to developing universal humanoid robots, recently raised a whopping $675 million in financing from giants such as Nvidia and OpenAI, with its valuation soaring to $2.6 billion. It is not difficult to see that in the emerging field of robot AI, a battle among giants has already begun.

The technical route is controversial

Although Pi's ambitious goals and financial strength are impressive, its technical approach is also controversial.

According to the disclosed information, Pi plans to purchase various off-the-shelf robot hardware platforms and then train its own AI models on them, rather than developing its own robot hardware. There are big questions as to whether this "buy ready-made unified standards" approach can ultimately achieve universal intelligence. After all, robot hardware varies greatly, and unified training may not be able to fully tap the performance potential of each platform.

In contrast, some competitors such as Figure AI choose to independently develop robot hardware and combine it with software. This "hardware and software integration" route may be more conducive to the ultimate realization of general intelligence, but the cost and cycle are greatly increased.

Another controversial issue worth noting is Pi's decision to temporarily put aside its focus on developing humanoid robots. In contrast, competitors such as Figure are betting big on making robots act and think like humans. Whether it is necessary to pursue anthropomorphism is still undecided.

Lachy Groom, one of the co-founders of Pi, once said, "I think what people build with humanoid robots is really cool, but fundamentally, it's the brain, not our hardware, that makes humans interesting." They believe that teaching robots of any shape and structure to learn is the solution to the fundamental problem.

Last words

In the field of artificial intelligence, language models are undoubtedly the most dazzling stars in recent years. From ChatGPT to the current 10,000-model war, the amazing performance of large language models has shown the world the huge potential of artificial intelligence. However, to truly realize the dream of artificial general intelligence, the language world alone is far from enough. The physical world is the field that artificial intelligence needs to truly conquer.

The birth of Physical Intelligence sends a strong signal that AI is extending from the virtual world to the physical world. Perhaps this will become a milestone turning point, indicating that machines will control reality and bring a new technological revolution experience to mankind.

The road Pi is currently on is the final frontier of AI, a long and bumpy road. When will the robot system brain stir reality? We can only wait and see, and hope that this team of top talents can live up to expectations and write a strong stroke on the journey of artificial intelligence.