The company announced a significant increase in the production of its Figure 03 model, moving from one robot per day to one robot per hour. They believe the primary advantage is no longer mechanical prowess but the ability to rapidly collect data, train AI, and manage entire fleets of humanoid robots.
American firm Figure AI has reported a dramatic escalation in the manufacturing pace of its humanoid robots. According to the company, in less than 120 days, the output of the Figure 03 model has surged from one unit daily to one unit hourly, signifying a 24-fold enhancement in their assembly line’s productivity.
The developers have stated that over 350 third-generation robots have been produced to date, along with more than 9,000 actuators of various types incorporated into the humanoid designs.
While these volumes might seem modest in the automotive sector, for the humanoid robot market, this represents one of the most substantial production leaps in recent years. The majority of companies in this industry are still occupied with research, demonstrations, and limited pilot projects.
Figure suggests that the industry is entering a new developmental phase. Previously, the focus was on videos showcasing robots dancing, running, or carrying objects. Now, the competition is gradually shifting towards mass production, reliability, fleet management, and the accumulation of data for AI training.
The company explicitly links its production growth to advancements in artificial intelligence. Each operational robot amasses a vast quantity of information regarding movement, errors, its surroundings, and interactions with objects. This data is subsequently utilized to refine control algorithms and enhance autonomy.
This approach cultivates a positive feedback loop: the more robots operate in the real world, the more data the company acquires; as the AI improves, it becomes simpler to expand deployment; and with increased deployment, new data is gathered more rapidly.
To scale its production, Figure has developed proprietary software that orchestrates over 150 interconnected workstations. A key challenge has been combating manufacturing defects and improving the quality of components sourced from suppliers.
The company reports that its factory incorporates over 50 interim quality checks and more than 80 final tests for each robot. Additionally, battery testing and burn-in tests are performed, during which the machines execute thousands of repetitions of squats, steps, runs, and arm movements to identify any latent defects.
Figure claims that the success rate for initial battery checks has reached 99.3%, and the overall success rate for robot assembly now exceeds 80%.
Concurrently, the company is establishing the necessary infrastructure to manage large robot fleets. Systems for remote software updates, fault diagnostics, service management, and overall device fleet monitoring have been developed.
This aspect might prove as crucial as the robots themselves. While conventional industrial manipulators operate in highly controlled environments, humanoids must function in warehouses, offices, retail spaces, manufacturing facilities, and, in the future, even private residences. As the number of these machines grows, their management increasingly resembles the operation of cloud IT infrastructure.
In parallel, Figure has announced progress in its Helix AI platform. The new System 0 controller integrates data from cameras and the robot’s internal position sensors, enabling it to navigate stairs and uneven surfaces with greater confidence.
Of particular interest is the fact that training is conducted entirely in simulation using reinforcement learning techniques. Subsequently, the models are transferred to physical machines without requiring further fine-tuning. The challenge of transferring skills from the virtual to the physical realm is considered one of the most complex problems in modern robotics.
If these methods prove truly scalable, they could significantly accelerate the development of the entire industry. In such a scenario, the critical factor for success will no longer be the most impressive robot demonstrations, but rather the ability of companies to rapidly produce thousands of devices, maintain their operation, and continuously enhance their intelligence through real-world data.
About the Author
