Some time ago, OpenAI and Microsoft teamed up to drop two bombshells: the much-sought-after GPT-4 and the Microsoft 365 Copilot released with lightning speed. With the former's powerful graphic understanding and generation capabilities, the latter automates a lot of work, including writing documents, converting Word into PPT, generating charts based on Excel data...
For a while, the AI model seems to be able to do anything, but some tasks that require the body cannot be done for the time being. So, the pressure came to the robot.
What chemical reaction will AI large model plus robot produce? A recent Google study has given a preliminary answer: they trained a concrete multimodal language model with 562 billion parameters - PaLM-E, and then used this model to drive robots to complete tasks autonomously. In the process of performing these tasks, the robot must be able to see, "think" like a human, analyze the situation in front of it, and then develop an action plan and implement it.
For example, you can ask the robot directly, "If a robot wants to function here, what steps should it take?" PaLM-E can give the answer: first clean the table, remove the trash, then move the chair, wipe the chair, and finally put the chair back in its original place. For ordinary AI models, it is enough to answer these things. But the difference is that Google's big model has a body, so it can do all the work mentioned above. This gives everyone more room for imagination.
Of course, Google is not the only player in this field called "embodied intelligence". Many research institutions and start-ups at home and abroad have taken it as one of the directions of development. Among them, a robotics company called "1X" successfully attracted the attention of OpenAI.
Founded in 2014 as Halodi Robotics, 1X's goal is to create robots with practical value that can be applied in the real world to augment the global workforce.
Halodi Robotics released a demo of its own robot
In a recently announced Series A2 round, the company raised a total of $23.5 million. Importantly, the round was led by OpenAI Startup Fund, with participation from Tiger Global and a consortium of Norwegian investors such as Sandwater, Alliance Ventures and Skagerak Capital. With the funding, 1X plans to ramp up research and development of the bipedal robot model NEO, as well as mass production of its first commercial robot EVE in Norway and North America.
The robot model NEO displayed on the official website
Robotic EVE
On the official website of 1X, we can see the company's path planning for two major robot products: EVE has the ability to move gently, manipulate objects and interact with the world, which is very close to real-world applications; NEO's goal is to explore how artificial intelligence is formed in human-like bodies, which we call embodied AI. This may be both OpenAI's choice to invest in 1X and the reason why former Google robotics senior research scientist Eric Jang chose to join the company.
Eric Jang left Google at the end of March 2022 (after a six-year stay) and announced on April 25 that he would join 1X (then called Halodi Robotics) as vice president of AI.
Eric Jang was very careful in choosing his next home, and the most important deciding factor was whether the company had a technological advantage that was years ahead of its competitors. In his opinion, Halodi, which specializes in humanoid robots, has met this condition.
"The moat I personally bet on (by joining Halodi) is 'humanoid robots 5 years ahead of any other company'. Halodi already has it, and Tesla is developing their equivalent. My main work at Halodi started out as training models to solve specific customer problems in mobile operations, while also developing a roadmap for AGI: how to compress large amounts of embodied first-person data from humanoid form to produce general intelligence, theory of mind, and self-awareness." Eric Jang wrote in a blog post.
OpenAI's investment in 1X has put the company in the spotlight, and Brad Lightcap, an OpenAI startup fund, said they believe 1X could offer the approach and impact on the future of work. Griffin Schroeder, Partner at Tiger Global, also expressed enthusiasm for 1X's mission: "We believe that 1X's robots are revolutionizing the robotics space, and we are excited to invest with OpenAI to support their continued growth."
However, fusing AI with robots may not be as easy as it seems. Eric Jang also mentioned when he first joined the company, "Embodied AI and robotics research has lost some of its luster in recent years, as large language models can now interpret jokes, while robots are still picking and placing at unacceptable success rates." But he still chose to "bet". Because he believes that "it is not enough to train models in the bit world".
Eric Jang presents the current state of generalization research in the field of robotics in his blog. He mentioned, "Many robotics researchers are still training small models and have not used Vision Transformer!"
In the year since joining Halodi, Eric Jang has been closely following progress in the direction of AI-based models, trying to close the gap between robots and AI-generated models. In a blog titled "How We Make Robots More Like Generative Models," Eric Jang compares generative models and robotics across three different dimensions and thinks about how to better connect the two. The results may be reflected in the NEO robot that will be available this summer.