Figure says its Figure 03 humanoid robot arrived at Hall 52, an assembly and logistics hall at BMW Group Plant Spartanburg, on June 30, 2026. The company is demonstrating the robot performing automotive parts sequencing: selecting components from carts or bins and placing them in the correct slots for an assembly line.
The demonstration is significant because it combines several tasks that are often treated separately in industrial robotics. Figure 03 must identify parts whose positions and orientations can vary, manipulate thin-walled components, reposition its body, maintain balance and pull a large metal cart on caster wheels. Figure describes Helix 02, its “pixels-to-actions” vision-language-action system, as coordinating these actions.
The announcement is a company account rather than a peer-reviewed study or independently audited performance report. It does not establish how long the demonstration will run, how reliably the robot performs, or whether BMW has accepted it for sustained production use.
Contents
What changed at BMW
Figure presents Figure 03 as a move beyond the earlier Figure 02 task described in its BMW partnership. According to the company, Figure 02 was deployed on BMW’s assembly line in 2025 for sheet-metal pick-and-place work. Figure says that system contributed to the assembly of 30,000 cars in the previous year.
The new demonstration focuses on sequencing rather than continuously presenting identical components in a fixed orientation. In a sequencing workflow, the robot must select the required parts and supply them to the assembly line in the needed order.
That difference matters because parts in carts and bins may be shifted, rotated, partially obscured or otherwise positioned differently from one cycle to the next. Figure describes Figure 03 as the latest generation in this BMW work, but the announcement does not provide a full specification comparison between Figure 02 and Figure 03.
How the workflow works
Figure says Figure 03 uses Helix 02 to coordinate its hands, arms, torso and feet while it performs the task. The robot is described as manipulating parts with both hands, adjusting its foot placement, shifting its body to preserve reach and balance, and placing each part into the correct slot.
The system also pulls a large metal cart on caster wheels. This gives the demonstration two different physical requirements:
- precise manipulation of individual, thin-walled parts; and
- forceful whole-body movement while handling a heavy mobile object.
A vision-language-action, or VLA, system links visual observations and task-level instructions to robot actions. Figure describes Helix 02 specifically as a pixels-to-actions VLA. In this setting, that means the controller is intended to turn what the robot sees and what it has been instructed to do into coordinated movements.
The broader control problem is known as whole-body or loco-manipulation. It requires the robot to coordinate locomotion, balance, arms and hands instead of treating the robot as a stationary manipulator mounted in a fixed position. Figure presents Helix 02 as providing visual-motor control that continuously adapts movement and corrects small errors during the workflow.
Why parts sequencing is technically difficult
Conventional industrial automation is highly effective when the motion, object position and presentation are tightly controlled. A fixed robot can repeatedly perform the same operation with high consistency, but changing object locations and orientations generally require additional sensing, specialised tooling or a more carefully controlled workcell.
Parts sequencing introduces variation into that process. The robot must first determine what it is seeing and where the relevant component is located. It then has to grasp the part, maintain control while moving, travel or reposition itself if necessary, and place the component into the correct location.
The challenge is therefore not simply whether a humanoid robot can pick up an object. It is whether perception, dexterous manipulation, balance and movement can be combined into one reliable manufacturing workflow.
For manufacturers, this type of system could be relevant if it can work within an existing assembly and logistics environment without requiring the entire workcell to be redesigned. That remains a potential application rather than an established production advantage.