In two years since its founding, Figure AI has gone from a little-known startup to one of the most celebrated robotics companies. The company recently introduced Figure 02, the second generation of its humanoid robot.
From a hardware and software perspective, the company has achieved a lot in a short time. And it has also garnered the attention and backing of important names. In February, Figure raised $675 million from OpenAI, Microsoft, Jeff Bezos, and Nvidia at a $2.6 billion valuation.
However, the company does not have an easy road ahead, and it is up against powerful players as well as unknown challenges. In a recent interview with Peter Diamandis, Figure CEO and founder Brett Adcock shared a bit more about the company’s roadmap in the coming years. I strongly suggest watching the full interview, but here are some of my observations.
First of all, this is a very serious and dedicated team. They are not riding the wave of hype surrounding generative models. They have made a bet on the direction of progress around AI and robotics and are working very hard to achieve it.
In the interview, Adcock said that the company is currently manufacturing one robot per week. In an X post, Adcock added that they will be ramping production up to two robots per week in the next two months. At the same time, they are developing their production line to scale production next year.
This strategy of slowly increasing production is interesting from a few standpoints. It is clear that Figure 02 will not be the final iteration of the robot, and Brett explained in the interview that they will continue to improve the hardware until “software becomes an issue.” This means that they plan to create a generalized platform that can be used for various applications, which is the natural vision for humanoid robots.
However, to perfect the robot, they need not only to train its models but to perfect it in real-world settings. There is a lot of research and attention in using the latest foundation models in robotics, from large language models (LLM) to vision-language models (VLM) and vision-language-action (VLA) models). And we have seen impressive advances in models that work with different robot morphologies, models that can automatically create reward functions for robotic control, and models that can reason over instructions and map them directly to robot commands.
All of these techniques will take you much of the way. But getting over that last mile and creating robust real-world applications still requires difficult engineering in hardware and software. By continuing to assemble and deploy their robots, the Figure team will continue to improve the hardware without committing to a mass production infrastructure. At the same time, they will have a small fleet of robots that will gradually continue to grow.
How will they use these robots? They are already working with several partners in different industries, where they will deploy their robots to gather data to improve their models. As much as we have come to bridge the gap between simulation and reality, there is still no ultimate replacement for data gathered from the real world, and access to quality data will be one of the key differentiators between companies competing in the space.
Going back to this post from last month, this strategy will enable them to gradually scale this cycle that is at the heart of improving Figure’s robots.
Figure still has a long way to catch up with Tesla, which is arguably positioned to become the leader in humanoid robotics. Tesla has a great engineering team, immense financial resources, a great data infrastructure, and vast industrial settings where it can train and test its robots. But with its gradual scaling, Figure will be able to position itself as a serious competitor in the space. And per the interview with Diamandis, Brett believes that the market for humanoid robots will be so vast that no single company will dominate the industry.
And its partnerships with OpenAI, Microsoft, and Nvidia will help propel it forward. We don’t know much about the details of these partnerships aside from the fact that OpenAI is helping them fine-tune models for their robots and Nvidia is providing them with GPU resources. I’m interested to know how tasks and labor will be distributed between cloud-based and on-device models. I suppose that the OpenAI models will serve at the high-level planning and reasoning stage and the on-device VLMs will take care of the minute details of applying those plans in the real world. But these are just my speculations.
Another interesting aspect of the Figure team is its organizational structure. Brett has made it clear that they will be an engineering-first team for the time being. They have raised a big chunk of money, but he is very cognizant of how fast that money can be wasted if not managed properly. He has traded the big board of directors and high-paying C-suite executives and product managers with an engineering-focused team.
According to the interview, they are a team of around 100 engineers, horizontal, nimble, iterating fast to create the best humanoid robot as fast as possible. Down the line, when they reach the state where they have a robust platform, they can bring in a business team that can start productizing the platform. I’m seeing a lot of parallels between the way Brett approaches humanoid robots and Elon Musk’s early years at SpaceX.
Brett has a vision of deploying thousands and millions of humanoid robots that cost lower than cars to manufacture. It is a very ambitious goal and might not come to fruition as he envisions it. But we have seen time and again that those who shoot for the stars often create long-lasting positive effects, even if they don’t reach their destination.