news

On August 2, Bloomberg reported that Covariant, an American AI software company focusing on embodied intelligence, recently received acquisition intentions from Amazon. Currently, neither Amazon nor Covariant has responded to the news.

Covariant completed a $75 million Series C financing in April 2023, bringing the total financing to $222 million. After the completion of this round of financing, Covariant's valuation reached $625 million. Covariant also has a luxurious investment team, including Geoffrey Hinton, Yann LeCun, Fei-Fei Li, Jeff Dean, Bill Gates and other big names, as well as investment institutions such as Index Ventures and Radical Ventures.

Covariant's core business focuses on the "Covariant Brain" software system and its AI robot solutions.The warehousing scenario is the scenario where it chooses to be implemented first.In March this year, Covariant also launched the first commercial robot foundation model RFM-1 (Robot Foundation Model-1). This model can help warehouse robots understand and perform related tasks more quickly through natural language input such as videos, pictures and texts.

At present, Europe's largest e-commerce group Otto Group, logistics company Radial and pharmaceutical distributor McKesson Corp are also major customers of Covariant.

In fact, the news of this acquisition is not unexpected. Amazon has been a leader in warehouse automation since it acquired Kiva Systems in 2012. In recent years, Amazon has successively acquired companies related to warehouse robots such as Cloostermans and Canvas Technology.

“Amazon can manage all the automation equipment through one platform.”Brittain Ladd, a logistics industry consultant and former Amazon executive, revealed that Covariant is an attractive acquisition target for Amazon because its software platform can control a variety of robotic hardware, which is equivalent to creating a "universal brain" for multiple hardware.

But some investors told Silicon Rabbit that if Covariant is acquired by Amazon,This may not be a good signal for the embodied intelligence track. Because if Covariant chooses to be acquired by Amazon now between independent development and listing and being acquired, it may mean that Covariant lacks confidence in commercialization, and also means that the prospects for the development of embodied intelligence are difficult.“Overall, Covariant is likely to accept Amazon’s acquisition invitation.Not too high。”

Three former OpenAI employees join forces to start a business

Aiming at the direction of warehouse robots

One of the important reasons why Covariant is popular among investors is thatFounding Team。

First is Andrew Ng's founding disciple,Pieter AbbeelHe is the first doctoral student of Andrew Ng, former chief scientist of Baidu and AI expert, and mainly focuses on research areas such as robotics and robot learning.

He also founded the Berkeley Robot Learning Lab and is now a tenured professor at Berkeley University. Currently, Pieter Abbeel serves as the chief scientist and co-founder of Covariant.

Pieter Abbeel

The second largest group is the Chinese born in the 1990sPeter Chen, currently serves as CEO of Covariant. He is one of Pieter Abbeel's students, studying for a bachelor's, master's and doctorate degree at Berkeley, and also interned at Meta.

According to relevant media reports, an Index Venture investor once commented on Chen Xi: "Chen Xi is not only a smart AI technology expert, but also an excellent businessman. (He) has strong control over market strategy, cultural construction, talent recruitment, business negotiations and other aspects."

Group photo of Covariant founders and team

As Covariant’s CTO,Rocky Duan) not only served as CTO of the auction website Sellegit during his undergraduate studies, but also completed his doctoral studies at the University of California, Berkeley in two years, and co-founded Covariant during his doctoral studies.

The above three have worked in star AI companiesOpenAIAfter working for a while, they resigned from OpenAI in 2017. At this time, the media broke the news that OpenAI's robot team was disbanded. Therefore, many people jokingly called Covariant the continuation of OpenAI's robot dream.

And the last co-founder,Tianhao ZhangHe once interned at Microsoft and started pursuing a PhD at Berkeley in 2016.

It is reported that Covariant's entrepreneurship originated from a gathering between Chen Xi and Duan Yan.

At that time, their team had just published a paper on reinforcement learning, and Chen Xi believed that reinforcement learning would narrow the gap between AI robots and humans, but it lacked some implementation directions, such as what objects the robot needed to manipulate and with what accuracy.

Chen Xi realized that there was a gap between academic research and industrial applications, which was one of the reasons why they founded Covariant. Covariant hopes to use the best AI scientists to solve practical problems. Based on market demand, they eventually set their sights on warehousing and logistics.

To build the robot's "brain"

Covariant Brain has an accuracy rate of up to 99%.

“We only sell AI software.”Chen Xi had a clear positioning for Covariant from the very beginning.

Currently, Covariant's main method of monetization is to sellCovariant BrainAlthough Covariant launched its first commercial robot basic model in March this year,RFM-1, but it is not clear whether RFM-1 has been commercialized.

What kind of software would make the "Father of Deep Learning" Geoffrey Hinton express his approval and even tweet to express his regret for not investing enough?

Covariant Brain is amazing in two ways:High-quality training data and generalization capabilitiesThis means that when a robot faces a new task or a new object, it only takes a short time or a few learnings to master new skills. In most current robotic solutions, robots need to be reprogrammed or take a long time to master new skills.

One of the biggest challenges facing warehouse robots in work scenarios is how the robots can adapt to complex work scenarios and accurately identify thousands of product categories and grasp them precisely.

Covariant said in 2020 that its Covariant Brain system can pick up and pack about 10,000 items with 99% accuracy. One of the secrets of its success is that the data used to train Covariant Brain is generated by real-time interaction between robots and the real world. The data is more valuable and more accurate, and the training effect is better.

"To enable robots to perform tasks autonomously, you need high-quality data. And if you need high-quality data, you need to collect it purposefully," Chen Xi said in an interview with the media. Covariant has always focused on "special scenarios", which are difficult to see in laboratory environments but exist in large numbers in the real world, making data more difficult to obtain.

Figure 1: Warehouse robots face various complex objects and special events

Currently, when robots are faced with ever-changing goods and tasks, Covariant Brain helps them autonomously identify and adjust their movements to complete grasping tasks, which is also the key to improving logistics efficiency in logistics systems.

Similarly, Chen Xi believes that AI robots need to have generalization capabilities. "If AI robots need multiple human interventions to pick products, then AI robots will not be valuable," he said.

To achieve generalization, Covariant believes that as long as enough real robot interaction data is used to train the model, a model that accurately simulates changes in the physical world will emerge. Covariant uses data to verify that the Scaling Law is applicable to the field of robot learning. The more training data there is, the fewer times a grasping failure needs to be repeated by 43%, and the robot's grasping speed is getting higher and higher.

Image source: Covariant Blog

However, the price of Covariant Brain is not cheap for customers. One Covariant customer once revealed: "If the price of hiring a sorting worker is $40,000, the price of Covariant Brain is about $30,000."

Want to be the GPT of the robot world

Covariant Launches RFM-1

On March 11 this year, Covariant launched its first Robot Foundation Model RFM-1 (Robot Foundation Model-1). The model is based on the Transformer architecture, has a parameter scale of 8 billion, and is trained using Internet data + real robot operation data.

RFM-1 supports the input of multimodal data, such as text, robot movements, sensor data, pictures, videos, etc., and robot operation instructions can be input through natural language.

Another notable feature of RFM-1 is that customers do not need to know how to program, and can use natural language to perform human-machine collaboration. It allows customers to use concise text language to guide the robot to perform tasks, and also caters to the ever-changing business needs of each customer and adapts to some special scenarios.

RFM-1 can also proactively ask humans for help. For example, if it encounters difficulty sorting a specific item, the robot can communicate and explain the difficulty to the operator. The operator can then teach the robot a new sorting strategy, and after repeated repetitions, the robot can learn the new strategy and apply it in future operations.

All of these features are possible thanks to the massive world model that Covariant built for RFM-1.

Luxury investment team

Gaining attention from industry, academia and research

Today, there aren’t many embodied intelligence startups that can match Covariant’s portfolio.Almost all of the AI ​​academic giants, well-known investment institutions, and Internet giants are among Covariant’s investors.

Data source: Crunchbase

Geoffrey Hinton, Turing Award winner and one of the three giants of deep learning, once wrote an article specifically complaining that he regretted investing too little in Covariant. If he were given another chance, he would reinvest a hundred times the previous amount.

In addition to Geoffrey Hinton, other AI academic giants Jeff Dean, Fei-Fei Li, and Yann LeCun all participated in the investment in Covariant.

In response to Hinton, Chen Xi said, “The reason why Hinton is optimistic about us and invests in us is that, on the one hand, he knows many people in our team and recognizes our strength; on the other hand, he has already seen and affirmed Covariant’s technical achievements.”

In addition, many well-known investment institutions have also expressed their optimism about Covariant. Index Ventures and Radical Ventures participated in the lead investment in Covariant's Series C financing. Some technology giants are also hidden behind some investment institutions, such as Baidu Ventures and Samsung NEXT.

This year, a large amount of capital and outstanding entrepreneurs have poured into the embodied intelligence track in North America. Skiled AI, founded by two CMU professors, is valued at 1.5 billion US dollars. Professor Fei-Fei Li started a business and the valuation of her company reached 1 billion US dollars at the beginning. Physical Intelligence, founded by a team of Stanford professors, is valued at 400 million US dollars. They all want to become the OpenAI in the field of robotics and aim at the GPT model of robots.

However, judging from the current corporate ability to attract capital and valuation,The early development of embodied intelligence cannot be separated from a large amount of financial support, and because of the hardware costs, its commercialization may be more difficult than large models.

▼

Be acquired or develop independently?

Covariant's choice is worth looking forward to

In 2012, Amazon purchased Kiva warehouse robots for $775 million, which caused a sensation in the industry, and logistics warehouse automation began to flourish. Although 12 years have passed, more than 80% of warehouses still have not deployed automation solutions. In today's highly developed e-commerce, logistics warehouse automation is still one of the key breakthrough directions for startups.

Judging from the configuration of the founding team, the lineup of investment institutions and the business progress, Covariant can be regarded as a leading company in the field of embodied intelligence.

By choosing to join Amazon, Covariant can naturally enjoy the shade of a big tree, without having to worry about capital issues, and also have commercial application scenarios within reach. Naturally, Amazon's acquisition of Covariant has another story to tell, and the boost to its stock price may easily exceed the price paid for it.

But will Covariant be willing to do so?

According to PitchBook data, Covariant was valued at $625 million in a round of financing in April 2023. Given that Covariant's public business progress information is ahead of Skiled AI and the current market enthusiasm for AI+robotics, it is unknown whether Amazon can offer an acquisition price of more than $1.5 billion.

But even so, referring to the valuations of OpenAI and Anthropic, if Covariant insists on independent development, can it be valued at tens of billions or even hundreds of billions of US dollars like the former two?

Interaction at the end of the article:

What do you think of Amazon's offer for Covariant?

Tell us what you think in the comments section~