Open AI Robots: Rethinking the Future of Humanoid Robotics

Big Data Digest authorized reprint from Robot Lecture Hall

Author: Li Xin

On the night of March 13, Figure AI showcased its first research results in collaboration with OpenAI. In the released video, we can see that the Figure 01 robot can engage in real-time conversations with humans and perform complex hand grasping and placing operations.

The development progress of Figure 01 is remarkable. It completed three rounds of financing in less than two years, with a total market value exceeding $2 billion. Founder Brett Adcock has been referred to as the “Elon Musk 2.0.” Previously, he gathered key figures from Boston Dynamics, Tesla, Google DeepMind, and Archer Aviation to create the world’s first commercially viable general-purpose humanoid robot.

Open AI Robots: Rethinking the Future of Humanoid Robotics

The biggest technological route of Figure 01 lies in its reinforcement learning capability. In February of this year, Figure AI just reached a cooperation with OpenAI to enhance Figure 01’s intelligent interaction capabilities through GPT 5.0.

Elon Musk tears up the first page of Open AI’s lawsuit

This update of Figure 01 has put significant pressure on Tesla’s Optimus. It is noteworthy that Elon Musk once tore up Open AI and filed a lawsuit against it, claiming it violated the original intention of being “non-profit” at the establishment of the company. It seems unlikely that Optimus will integrate with Open AI in the future. As of the time of writing (March 4, 10:44 AM), Elon Musk has not commented on Figure AI’s latest developments.

Two years, three rounds of financing, Figure AI’s valuation has reached $2 billion

Figure AI was founded in 2022 and is headquartered in California, USA. It is a startup dedicated to developing autonomous general-purpose humanoid robots. The company has the Figure 01 bipedal humanoid robot product, which is approximately 170cm tall, weighs 60KG, has 41 degrees of freedom, can effectively carry 20KG, moves at a speed of 1.2 meters/second, and has a battery life of up to 5 hours. The special feature of Figure 01 is its strong autonomous learning ability. It is entirely AI-driven, and after integrating AI, Figure 01 can complete specified tasks without pre-set programs and possesses a certain degree of generalization learning ability.

In October 2023, Figure AI released its first Figure 01 video on YouTube. From the video, it can be seen that the robot can already stand and perform basic autonomous tasks. By the end of the year, Figure 01, through reinforcement learning algorithms, will have preliminary observation and learning capabilities. In January 2024, Figure 01 will be able to operate independently to brew coffee after observing human actions and learning from several hours of video, even correcting the position of the coffee capsule autonomously.

In terms of commercial landing, five months after the first exposure of the Figure 01 robot, it has collaborated with BMW Spartanburg’s factory for the first humanoid robot pilot. The robot will start with “a small amount of work” and, over time, these AI-driven humanoid robots will take on more manufacturing tasks, including box handling, picking and placing tasks, and pallet loading. After the assessment period, in the future 12-24 months, Figure 01 will be integrated into BMW’s manufacturing processes, including body workshops, sheet metal, and warehouses.

In February of this year, Figure AI announced the completion of a new round of financing of approximately $675 million (about 4.87 billion RMB) to develop humanoid robots that provide labor for repetitive and dangerous warehouse and retail tasks. Silicon Valley tech giants such as Amazon founder Bezos, Nvidia, Microsoft, and OpenAI provided financing support for Figure AI. Before financing, Figure AI’s valuation had already reached $2 billion.

Figure 01’s ambition is to train a world model to control billion-unit-level humanoid robots

The operational logic of Figure 01 is as follows: testers input the images captured by the robot’s camera and the text information recorded by the onboard microphone into a multimodal model (VLM) trained by OpenAI that understands images and text. This model processes the entire history of the dialogue, derives a language response, and then replies to humans through text-to-speech.

Figure 01 integrates its camera to send data to the large visual language model trained by Open AI, while Figure’s own neural network also “collects images at a speed of 10 Hz through the camera on the robot.” In addition, Open AI is also responsible for understanding spoken language, and all this incoming information is transformed by Figure’s neural network into “fast, low-level, agile robot movements.”

During the demonstration, Figure 01 was not operated remotely; the video was filmed at actual speed. Founder Brett Adcock stated, “Our goal is to train a world model to operate billion-unit-level humanoid robots.

Humanoid Robots: The Next Phase of Commercialization?Computing Power?What pressures and challenges do we face?

Humanoid robots have entered the second half of the competition, where the key to the competition no longer focuses on joint motors, reducers, and controllers, as many domestic humanoid robot companies can already match the leading companies in Europe and America in single indicators. What is truly lacking domestically is exploration in commercial landing and computing power.

In 2023, Agility Robot’s Digit has been deployed at Amazon’s Sumner BFI1 operations center for factory testing. Figure 01 has partnered with BMW Spartanburg’s factory for the first humanoid robot pilot. Apptronik’s Apollo robot is heading to NASA for special tasks. Although domestic companies such as UBTECH and LEJU have also entered commercial exploration, they are clearly lagging behind European and American companies in terms of timing. It can be said that commercial landing exploration is continuously correcting technological deviations and racing against time. In the future, whoever can first achieve mass production will be the first to reap the dividends of the trillion-dollar market of humanoid robots.

Computing power is an uncontrollable factor. Due to the impact of the global economic landscape and trade barriers, domestic humanoid robot companies find it difficult to cooperate directly with Open AI, so the burden falls on domestic large model companies. As of now, there are more than 168 domestic companies leading the development of AI large models (data source: https://github.com/wgwang/awesome-LLMs-In-China), of which only 24 are conducting research on general models, 40 are for research/medical/government/public services, and the rest are for commercial applications and services. Despite the large number of R&D companies, no competitor has emerged that can truly match Open AI. Large model computing power has become a bottleneck for the second half of China’s humanoid robots.

Conclusion and Future

In 2024, the singularity of the humanoid robot industry will arrive, and the market will show a more pronounced Matthew effect. The key factors determining the landing of humanoid robots in the future are the cost of robots and the degree of intelligence. We have a complete industrial system and manufacturing advantages in the supply chain. However, we are significantly lagging behind the United States in terms of model computing power, talent, and data reserves, making it challenging to overtake. In the second half of humanoid robots, as AI grows at an exponential iteration speed, it seems that we have little time left to think.

Rent! GPU Computing Power

A new batch of 4090/A800/H800/H100 is online

Especially suitable for enterprise-level applications

Scan to learn more ☝

Those who click “View” all become better looking!

Leave a Comment Cancel reply