When Figure AI published its retrospective on deploying its F.02 humanoid robot at BMW’s Spartanburg plant, the numbers looked impressive: 30,000 cars, 90,000 parts loaded, 200 miles walked. But what do those numbers actually tell us about where humanoid robotics stands today? Over an 11-month programme, Figure’s robot ran 10-hour shifts Monday through Friday on a single sheet-metal loading task, picking parts from racks and placing them onto welding fixtures. The company defined three KPIs: a cycle time of 84 seconds, placement accuracy above 99% per shift, and zero human interventions per shift.
The 30,000 cars figure
Figure 02 did not build those 30,000 cars. It loaded sheet-metal parts for one sub-process on one line. Dozens of human workers and conventional industrial robots handled everything else. At 90,000 parts over 1,250 hours, the robot loaded roughly 72 parts per hour: one every 50 seconds. But the task involves loading three parts per cycle, meaning one cycle took approximately 2.5 minutes. That is slower than the 84-second target cycle time Figure itself set. A trained human associate on a comparable BMW pick-and-place line would complete a full cycle in 30 to 60 seconds, well within the takt time the line was designed around. The 84-second target already builds in generous extra time for the robot. For now, humans would still be faster.
A claim in humanoid robotics is that robots can work around the clock, unlike their human counterparts. Figure 02 did not do this. It ran the same Monday-to-Friday, 10-hour shift pattern as human workers, maybe because the robot required human supervision, maintenance, and intervention handling outside those hours. Over a six-month operational window, a genuinely autonomous round-the-clock robot would have logged upwards of 4,000 hours. Figure 02 logged 1,250, roughly 29% of theoretical continuous capacity. One of Figure’s three KPIs was zero human interventions per shift. The company does not report whether this was achieved.
Hardware reliability
The deployment post also mentions hardware failures. The robot’s forearm was identified as the top failure point, a subsystem that required a complete architectural redesign for the next-generation F.03. Conventional industrial robot arms from manufacturers like FANUC or KUKA routinely accumulate tens of thousands of hours between failures in automotive environments.
Getting any humanoid robot onto a live automotive assembly line, without a safety incident, for six months, is a first. The task required real-time adaptive locomotion and 5mm placement precision under industrial conditions. And the engineering learnings that fed into F.03 represent the actual value of the exercise: not productivity, but knowledge.
This deployment is therefore better understood as a long-duration field test than a proof of industrial competitiveness. Figure 02 performed a single constrained task, at below-human speed, on a generous cycle-time budget, without running outside human shift hours, with unreported intervention rates. The robot’s performance the quantitative KPI’s fell short of what a trained human worker delivers on the same line. But Figure is clearly learning fast. Figure 02 has now been retired and Figure 03 is in development. The company says everything it learned at Spartanburg is baked into the new design.
