Robot Even Making 'Neoseure'... The Secret Behind Its Dramatic Change in Two Weeks [AI Bite News]

OpenAI Brain-Equipped Humanoid 'Figure01'
Understands Human Intent and Acts... "I Did Pretty Well, Right?"

The humanoid robot 'Figure01' has skyrocketed to stardom with just a 2-minute and 34-second video. Not only are its movements as natural as if enhanced by computer graphics (CG), but it also demonstrates the ability to think and act on its own. This result, released just two weeks after humanoid startup FigureAI announced its collaboration with ChatGPT developer OpenAI, has been met with amazement.

Let's look at the action of handing over an apple. Although it seems like a simple motion, the robot must go through many processes. First, to hear and understand the question "Is there anything to eat?" it needs language capability. It also requires cognitive ability to look at the table and distinguish what objects are present. Then, it must judge which item among the cup, plate, and apple is edible. The physical action of picking up the apple and naturally handing it to a person is yet another domain.

The robot’s ability to infer the next action and evaluate its own behavior is also impressive. When asked, "Where should the plate in front be moved?" Figure01 responded, "I should put it in the dish rack next to me," and acted accordingly. When the experimenter spilled trash on the table, the robot picked it up even without being instructed. When asked, "How do you evaluate your own actions?" it jokingly replied, "I think I did quite well." Its tone was also natural, not a stiff mechanical voice, even showing some hesitation.

Compared to when Figure01 was first unveiled in March last year, this is a remarkable advancement. At that time, Figure01 performed relatively simple tasks such as walking on two legs and moving boxes onto a conveyor belt. Then, last month, it unexpectedly received about 900 billion KRW in investment from OpenAI, Microsoft (MS), and Nvidia. It was decided that FigureAI would develop next-generation AI robots with OpenAI and receive infrastructure support from MS.

Equipped with a 'Smart Brain' through Collaboration with OpenAI... The Power of Large Language Models (LLM)

[Image source=AP Yonhap News]

The reason Figure01, which previously only performed simple actions, has grown rapidly lies here. By integrating OpenAI’s AI model, it gained a powerful brain. The fact that the robot is equipped with a large language model (LLM) means more than just improved performance. It changes the very way robots are trained.

Previously, teaching robots took a long time. To assign a specific task, every single action had to be programmed. For example, if the robot had to clean up spilled cola, it needed to be trained step-by-step: recognizing that the spilled cola is the target to clean, picking up the cola can, finding something to wipe the cola, and so on. If the spilled liquid was water instead of cola, or if it was in a glass instead of a can, the robot had to be taught separately.

With an LLM, the story changes. By combining vast textual information with the robot’s visual input, it can learn and execute skills on its own. Just as an LLM-based ChatGPT doesn’t need to be separately taught how to write emails or novels, an LLM-based robot can decide the tools or actions needed for new tasks. FigureAI stated, "Figure01’s movements are based on the robot’s own observation and judgment," adding, "It plans its next actions, reflects on memories, and can logically explain its thoughts."

Google’s AI robot 'RT-2,' unveiled in July last year, is similar. RT-2 flawlessly executed complex commands such as placing a Volkswagen toy car in front of the German flag, which was spread out alongside flags of the U.S. and Germany. Because it acquired knowledge from massive data, it knows what the German flag and Volkswagen car are without separate training. Hongseop Choi, head of the AI Technology Division at MaumAI, said, "Thanks to foundation models, companies that excel in AI can also excel in robotics," and predicted, "The speed of robot capability development will accelerate."

※Terminology Explanation: Foundation Models

AI trained on broad data sets capable of performing a wide range of tasks, not limited to specific cases