news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

Has the robotics industry’s “GPT-3” moment arrived?

Author | Xinxin
edit| Jingyu

Embodied intelligence, or humanoid robots, has now become the hottest investment project after AI. Whether it is OpenAI, Nvidia or Microsoft, they are all investing heavily in humanoid robot teams.

Now, a seed player is gaining the favor of the giants.

Recently, a company called Skild AI announced the completion of a $300 million Series A financing.Investors include Jeff Bezos, Japan's SoftBank Group, Sequoia Capital and Carnegie Mellon University, bringing the company's valuation to $1.5 billion。

The company was founded less than a year ago by two university professors in the field of robotics. Its team is building a "scalable robotics base model" as a universal "brain" for various types of robots and various real-world application scenarios, while also building systems that can be retrofitted to robotic hardware.

"Our long-term goal is to develop artificial general intelligence (AGI) grounded in the physical world, challenging the popular notion that AGI can only arise from digital knowledge," the company said.

Investors quickly poured money into this, believing that the "GPT-3 moment is coming" for the robotics industry。

What did Skild AI do right? Can it realize the ambition of AI + embodied intelligence?

01

"Large Model of Robot"

Founded in May 2023 by Carnegie Mellon University professors Deepak Pathak and Abhinav Gupta, Skild AI is developing intelligent systems based on the physical world and building basic models of robots - which can be understood as the "robot brain."

What is special about what they are doing? Traditional robotics focuses on collecting specific data to train robots to complete specific tasks, while Deepak and Abhinav use large-scale data to build a basic model through an adaptive architecture based on Transformer. They want to create a general, robust and emergent robot model.

The company said,It is breaking through the data barrier of robots. The amount of data for its training model is "more than a thousand times that of its competitors' models"Unlike robots that are designed vertically for specific applications, Skild’s model serves as a “general” brain for a variety of robot forms, scenarios, and tasks, covering functions such as manipulation, movement, and navigation.

The company's models are said to have real-world applications ranging from "quadruped robots" that are resilient in harsh physics to "humanoid robots" that can perform complex household and industrial tasks.

In theory, then, this “robot brain” could power Boston Dynamics’ quadruped robot “Spot” and Agility Robotics’ humanoid robot “Digit,” although the company has yet to announce specific partners.

Building a “universal” robot base model | Image source: Skild AI

Skild AI says its mission is to "revolutionize the future of physical work by developing the first truly intelligent physical systems, aiming to increase productivity and tap into human potential." Its vision is to "build general artificial intelligence (AGI) that is rooted in the physical world."

Although many people around the world are worried about AI or robots taking their jobs, companies that make robots generally like to say that they want to solve the so-called "labor shortage problem" or, more loftily, "liberate humanity."

Skild AI is clearly no exception.

They emphasized that the United States, for example, is currently facing a serious labor shortage, with 1.7 million more vacant jobs than unemployed people. Industries such as healthcare, construction, warehousing and manufacturing are the most affected, with an estimated 2.1 million manufacturing job vacancies by 2030.

Additionally, many of these jobs can be dangerous for humans, such as on oil rigs and in machine rooms, and Skild’s model enables robots to adapt to perform new tasks in dangerous environments, rather than having humans perform those tasks.

Abhinav Gupta, co-founder of Skild AI, said, “Universal robots can safely perform any automated task in any environment and in any type of implementation. We can expand the capabilities of robots, reduce their costs, and support the severely understaffed labor market.”

Skild AI claims the brain can adapt to a variety of hardware and tasks | Skild AI

People in the robotics industry often say that "robotics is hard", which has almost become one of the unsung laws of nature governing the field.

Moreover, many people think that robots are a hardware problem, but the founder of Skild AI believes that it is a software problem.。

Skild AI emphasizes that “scale is key” and says they are inventing cutting-edge machine learning algorithms, “focusing on leveraging the power of scale to provide unparalleled robustness in any environment. From construction sites to factories to homes, Skild Brain can adapt to unstructured environments like humans do.”

The “Skild Brain” is the so-called brain, which is said to be the “first scalable” robot foundation model that can adapt to different hardware and tasks and is “robust in model design.”

The company also revealed a Skild AI mobile control platform, powered by Skild Brain, on which "advanced AI algorithms and applications for robots can be developed", and they want to "make robot operation as simple as calling an API". Currently, the platform is not open, and only interested developers are allowed to register and join the early access waiting list.

In addition, they launched a security/inspection robotics platform that they claim provides solutions for automating visual inspection, data collection or patrol tasks.

Skild AI’s founding team | Image source: Skild AI

In July this year, Skild AI announced the completion of a $300 million Series A financing round led by Lightspeed Venture, Coatue, SoftBank Group and Jeff Bezos (through Bezos Expeditions), with participation from Felicis Ventures, Sequoia Capital, Menlo Ventures, General Catalyst, CRV, Amazon, SV Angel and Carnegie Mellon University. This financing brought the company's valuation to $1.5 billion.

Skild AI said it will use the new funds to improve its AI models while pursuing commercial deployment. The long-term goal is to create an AGI that "has human-like capabilities" and is "rooted in the physical world."

"We believe Skild AI represents a turning point in the way robotics is scaled, with the potential to transform the entire real economy," said Deepak Pathak, CEO and co-founder of Skild AI.

02

Indian RobotEntrepreneurship

According to the company website, Skild AI is hiring for multiple engineering positions. The current team includes members from Meta, Tesla, Nvidia, Amazon, and Google backgrounds, as well as students from Carnegie Mellon University.

The two co-founders, Deepak and Abhinav, were professors at Carnegie Mellon University. They have many years of research experience in robotics and AI and are well-known for their research in self-supervised robotics, curiosity-driven agents, and adaptive robot learning.

If we want to quantify their academic level, the two of them currently have an H-index of 150+ and more than 90,000 citations.

Skild AI co-founders Deepak (left) and Abhinav (right) | Image source: Skild AI

Deepak hails from a small town in India. While his peers moved to big cities to prepare for national exams, he stayed in his small town and still made it to the Indian Institute of Technology (ITT), the country’s top institution of learning, an achievement that made local headlines.

It is said that when he was in India, due to lack of conditions, he used to hand-write codes on paper at home and check them, and then run his programs during the limited Internet time in the local cafe. Later, this "small town legend" in India went to the United States to study for a doctorate, joined Facebook AI Research Institute (FAIR) to do research, founded a startup company that was acquired, and later chose to become a professor.

Abhinav is a tenured professor at Carnegie Mellon University. Like Deepak, he is of Indian descent and was a founding member and research leader of the FAIR Robotics Research Group.He and Deepak discussed the possibility of starting a business in ten years. In early 2023, they saw the acceleration of technological progress in their field and realized that it was time to make a big move.

So what exactly are the opportunities they see? In the pursuit of building general intelligence for robots, a key challenge has always been how to build a large model without large-scale data.

Unlike large language models, there is no readily available internet data in the robotics field. Therefore, they explored different strategies for learning from existing resources: online videos, teleoperation, real-world data, simulations, etc.

In 2015, they achieved the first 1,000-fold expansion of robot data, and in the following years, they experimented with human teleoperation and low-cost robot teleoperation platforms. In 2017, they proposed the famous curiosity-driven learning algorithm for building intelligent agents that can explore and learn autonomously. In 2021 and 2022, they broke through again, using the large-scale adaptive SIM2REAL (virtual to real world training) strategy and won the Best Robot System Award at the Robot Learning Conference.

These achievements lay the foundation for Skild AI's goal: a general model that can complete any task in any environment without specific training. Some investors believe that if Deepak and Abhinav can achieve this goal, they will achieve a breakthrough similar to GPT-3, and the results may be applicable to almost all fields.

03

Robotic

A GPT-3 moment?

Matthew Roberson, current director of the Robotics Institute at Carnegie Mellon University, endorsed them, saying, “Skild AI was founded by experts at the forefront of robotics innovation. I can’t wait to see how their cutting-edge technology revolutionizes the industry and continues Carnegie Mellon University’s long history of translational research.”

Other investors also praised Skild AI and seemed to be confident in its capabilities and prospects. Most of them were attracted by the contributions of the two founders to some of the world's most advanced robotics and AI laboratories.

Over the past few years, the internet has occasionally been struck by robots that can do extreme parkour, manipulate objects with their hands (including opening doors and drawers), climb stairs (both forward and backward, indoors and outdoors), and move naturally and smoothly, and these advances can allegedly be traced back to some of the two's academic work.

Some investors called them a "catalyst" for the advancement of robotics, believing that "their innovation in applying the core principles of the basic model to the real world has put the industry on the path to general robotics."

Some investors suggested that Skild AI "adopted a truly scalable approach" in building the basic model of robot operation and movement. "They revolutionized robotics technology from pre-programmed robots to dynamic adaptive robots, which has the potential to subvert the entire real economy."

Another investor, who has been investing in robotics companies for more than 15 years, called Skild the "most visionary" he had ever seen. "The models they are building will perform any task in any environment and on any hardware."

In the view of Felicis Ventures' investors, the race to develop general robotic intelligence has begun. No single idea can mobilize capital and talent as quickly as it does now. The general robotics basic model will become a key link in this chain.

The idea of ​​creating a "single model" that can reason, plan, and act in any environment and in any form is not new and has always been the holy grail that the robotics industry wants to seize. "The problem with this idea is that it was simply impossible to achieve a few years ago. Today, it may become a reality," said an investor from Felicis Ventures.

“When we first flew to Pittsburgh to see the Skild robot in action, we couldn’t believe our eyes. How could a startup achieve so much in such a short time?”

「The answer lies in scale. Large pre-trained Vision-Language-Action models (VLAs) exhibit the same emergent behavior as large pre-trained Language Models (LLMs), just as training an LLM on algebra makes it perform better on Spanish.Investors at Felicis Ventures believe that "Skild is taking this concept to the extreme."

What Skild is doing and wants to achieve is to use a thousand times more training data to enable robots to perform tasks they have never seen before. The investor said he has already seen signs of this.

Sequoia Capital investors also stated that they have "deep faith" in Skild AI's team, saying they "have the potential to achieve in the real world what OpenAI has achieved in the digital world."

One proof of their belief is that they gave money very quickly. Less than a week after meeting the founders, Sequoia Capital decided to invest in Skild AI, believing that this team was what they were looking for, saying that "GPT-3's moment is coming and will bring great changes to the world of robots, just like the progress we have seen in the world of digital intelligence."

These investors believe that while there is a lot of discussion about the impact of AI on human life, so far, much of the discussion is still focused on software, and there is a huge opportunity to integrate AI into robots, and AGI cannot be built only through digital knowledge. Although everything is in the early stages, robot manufacturers have the opportunity to develop smarter machines by leveraging advances in LLM, VLM, and code generation.

“Imagine a world where a single AI robotics base model can accomplish any task in any environment, on any robotic hardware. This would greatly expand the types of robots we can build, and the cost would be orders of magnitude lower than today.”

With AI in the software space becoming a game for the big guys, and the promise and appeal of integrating AI into robots.

A wave of AI hot money has shifted to the field of robotics.

*Header image source: Skild AI

This article is an original article from Geek Park. For reprinting, please contact Geek Jun on WeChat: geekparkGO

Geek Question

What do you think about AI+embodied intelligence?

Musk: I do have to be careful. Two people tried to assassinate me in the past seven months.

Like and followGeek Park Video Account，