Figure AI's F.03 returns to BMW’s factory floor

A Second-Generation Robot Takes on a Harder Job

Figure AI's newest humanoid, Figure 03, has taken up residence on BMW Group's Plant Spartanburg production line, replacing its predecessor Figure 02 in an ongoing deployment that has quietly become one of the more closely watched real-world tests of humanoid robotics in manufacturing. Figure 02 spent 2025 handling sheet-metal loading tasks that contributed to production of more than 30,000 BMW X3s. Its successor is now being asked to do something meaningfully more complex: sequencing.

Rather than simply moving pre-organized materials from point A to point B, Figure 03 is tasked with sorting unordered parts into the correct order and pulling loaded carts through the facility. This requires the kind of situational judgment and physical coordination that has historically been difficult for robots — recognizing parts regardless of orientation, deciding how to organize them, and physically maneuvering wheeled equipment through a dynamic factory environment.

Why Sequencing Is a Bigger Deal Than It Sounds

In manufacturing, sequencing is a logistics-heavy task that traditionally depends on human workers or rigid, purpose-built automation. Handing it to a general-purpose humanoid signals a shift in ambition: Figure isn't just deploying robots to do repetitive, fixed-path work, but is testing whether its machines can handle variability and decision-making that were previously the domain of human floor staff.

This matters because sequencing work is a plausible bridge task between narrow automation and more flexible robotic labor. If Figure 03 can reliably sort unordered parts and navigate carts without constant human correction, it suggests real progress toward robots that adapt to messy, real-world conditions rather than environments engineered around them.

The Role of Helix 02

Underpinning this capability is Figure's Helix 02 system, described as enabling "whole-body AI control." This points to the broader trend in embodied AI research: pairing large, generalized foundation models with physical robots so that perception, planning, and motor control are integrated rather than hand-coded task by task. Whole-body control suggests the robot is coordinating its arms, torso, and locomotion together to complete tasks, rather than relying on isolated, scripted movements for each subtask.

Why This Deployment Is Worth Watching

BMW's factory floor functions as a real-world proving ground rather than a lab demo, and continued iteration — replacing Figure 02 with Figure 03 within roughly a year — indicates a fast development cycle. If Figure 03 performs sequencing reliably at scale, it would be a meaningful data point for the broader humanoid robotics industry, reinforcing the idea that foundation-model-driven robots may be approaching genuine utility in unstructured industrial settings, not just controlled demonstrations.