After an 11 month pilot helped build roughly 30,000 X3 SUVs at BMW's South Carolina plant, the third generation Figure robot is taking on just in sequence parts work — a job requiring it to read mixed parts and match them to a specific build sequence, the real test of whether a…
BMW's Spartanburg, South Carolina plant has shifted the humanoid it has been testing since mid-2025 from one category of precision work to another. The robot that spent roughly eleven months inserting sheet-metal components into welding fixtures for the X3 is now being assigned to production logistics: taking bins of unsorted components and organizing them into sequencing trolleys that feed just-in-sequence delivery to the assembly line, with autonomous transport systems moving the trolleys between stations.
The transition was announced in March 2026 as part of BMW Group's broader Physical AI push. Spartanburg is one node in a multi-site, multi-vendor strategy that already includes a separate humanoid pilot in Leipzig, a new in-house AI hub, and a third program with robotics firm AEON. Within that footprint, it is where Figure AI is trying to graduate from a single body-shop task to a flexible plant-floor role.
The pilot that preceded the shift ran long enough to register a real production number. According to Figure AI, the prior-generation Figure 02 contributed to the production of 30,000 BMW X3 vehicles during its Spartanburg deployment, primarily by inserting sheet-metal components into welding fixtures. That figure is Figure AI's self-report, not an audited BMW production disclosure. BMW's PressClub release on the Figure 03 deployment cites no comparable vehicle count for the new work. The 30,000 X3 figure should be read as a vendor claim about the prior generation, not as evidence of what Figure 03 is now contributing.
What Figure 03 adds, on Figure AI's specification sheets, is a hardware refresh aimed at the new task profile. According to Figure AI's announcement, as covered by trade publication Interesting Engineering, the third-generation unit carries touch-sensing hands, palm-mounted cameras for fine manipulation, wireless charging, onboard AI speech, and softer exterior surfaces intended to make workplace contact less hazardous. Those are vendor specifications. No independent teardown or third-party benchmark has confirmed the tactile sensitivity, vision stack, or speech latency in a production setting.
The new task itself is the structural story. Body-shop welding-fixture insertion is repetitive and constrained: a known part, a known fixture, a known motion path. Sequencing unsorted components into the right trolley for the right build is a different problem. It requires the robot to read a bin of mixed parts, identify each one, and place it in an order that matches the specific vehicle moving down the line minutes later. If Figure 03 can hold that workflow under production tempo, the humanoid is no longer a single-task automation cell. It becomes a redeployable worker that can move between jobs without the fixed tooling changeovers that limit conventional robots.
That is the case Figure AI's founder and CEO Brett Adcock is making. "Humanoids are no longer lab experiments — they can be a valuable asset in establishing a flexible, reliable manufacturing workforce," Adcock said in remarks carried by trade media. The statement should be read as a vendor position rather than an independent assessment; Adcock's equity and role at Figure AI make him a stakeholder in the framing, not a neutral observer.
The honest counterweight is scope. This remains one automotive customer, in one plant, working with one humanoid vendor, on tasks that vendor hardware was designed to do. There is no public read on humanoid total cost of ownership in automotive logistics, no peer-OEM data on whether the sequencing workflow transfers to other assembly programs, and no independent benchmark comparing Figure 03 against fixed-automation alternatives for the same job. BMW's broader Physical AI footprint (Leipzig, the in-house AI hub, AEON) signals a multi-vendor strategy, but it also dilutes any single-vendor read of "humanoids work in automotive."
The watch items now are concrete. First, whether BMW publishes production data on Figure 03's sequencing output rather than letting Figure AI frame the numbers. Second, whether the sequencing task spreads beyond a single workcell; Figure 02 was, by Figure AI's account, a single body-shop operation, and the test of flexibility is whether Figure 03 takes on a second logistics role in the same plant. Third, whether BMW's other humanoid programs (Leipzig, AEON) report comparable throughput in their own task categories. The interesting question is not whether a humanoid can stand on a factory floor. It already can. It is whether one can move between production jobs without re-engineering the cell around it, and whether the OEM that owns the plant is willing to publish the data that would let an outsider answer that.