Figure at BMW: What Figure 02 Proved and Figure 03 Still Must Prove

The short answer

Figure 02 at BMW is one of the strongest public humanoid production cases because the customer confirms the site, task, duration, and vehicle-production context, while Figure publishes runtime and throughput. Figure 03 has not inherited that record. Its BMW evidence currently establishes a new logistics-sequencing demonstration and a planned deployment phase, not recurring production performance.

BMW says Figure 02 supported production of more than 30,000 BMW X3 vehicles over ten months at Plant Spartanburg. Figure reports more than 90,000 sheet-metal parts loaded, more than 1,250 runtime hours, and full deployment on every working day. The task was specific: load three sheet-metal parts into a welding fixture on an active automotive line.

The missing denominators still matter. Neither source publishes the number of robots, completed versus attempted cycles, achieved placement success by shift, intervention count, teleoperation, scheduled uptime, downtime, maintenance, safety events, or commercial terms. Figure discloses targets for cycle time, placement accuracy, and interventions, but it does not publish the achieved distributions.

Figure 03 is a separate generation and task. BMW says the robot will sequence unsorted components into a trolley for onward transport. Figure shows the robot picking thin-walled parts, adjusting its stance, and pulling a wheeled cart under Helix 02 control. Figure calls this the first demonstration of the workflow. That wording is evidence of demonstrated capability, not a completed pilot or production deployment.

The defensible classification is:

• Figure 02: customer-confirmed production deployment for a bounded automotive workflow, with substantial vendor-reported runtime and throughput.
• Figure 03: demonstrated at the same customer site for a new logistics workflow, with customer deployment intent but no published recurring operating record yet.

Evidence at a glance

Status reflects primary sources reviewed on September 17, 2026.

Field	Figure 02 sheet-metal loading	Figure 03 logistics sequencing
Customer and site	BMW Group Plant Spartanburg	BMW Group Plant Spartanburg, Hall 52
Workflow	Load three sheet-metal parts into a welding fixture	Sort unsorted components into a sequencing trolley and move it to a collection point
Evidence phase	Completed production deployment	First demonstrated workflow and announced next project
Duration	BMW says ten months; Figure describes an 11-month program and six months of daily runtime	Recurring duration not published
Schedule	Figure says a 10-hour shift, Monday through Friday, during full deployment	Schedule not published
Throughput	Figure reports 90,000+ parts; BMW confirms support for 30,000+ X3 vehicles	No recurring throughput published
Runtime	Figure reports 1,250+ hours	Not published
Task targets	84-second total cycle, above 99% placement success per shift, zero interventions per shift	No public acceptance thresholds
Achieved task quality	Aggregate parts and runtime published; target attainment not quantified	No repeated trial or shift result published
Human help	Figure says interventions were logged, but no count or rate is public	Teleoperation, interventions, resets, and supervision not disclosed
Hardware status	Figure 02 retired after the program	Figure 03 is the active successor generation
Strongest conclusion	Bounded production work occurred at meaningful duration and scale	A more complex sequencing behavior was demonstrated at the customer site

This comparison prevents three common errors: treating targets as achieved results, converting a demonstration into a deployment, and transferring a retired generation’s record to its successor.

What Figure 02 actually did

Figure 02 worked in BMW Plant Spartanburg’s body shop. An associate would ordinarily pick sheet-metal parts from racks or bins and place them on a welding fixture. Industrial robots then welded the assembly and fed it into the main line.

Figure describes a cycle that required three parts to be placed within a five-millimeter tolerance. It says the full cycle requirement was 84 seconds, including a 37-second loading phase, with individual placement in roughly two seconds. These numbers describe the process requirement and the performance target.

The production totals are more direct:

• more than 90,000 parts loaded;
• more than 1,250 runtime hours;
• operation during a 10-hour shift from Monday through Friday;
• more than 30,000 BMW X3 vehicles supported;
• an estimated 1.2 million robot steps, or more than 200 miles of movement.

BMW independently confirms the task and says Figure 02 supported production of more than 30,000 BMW X3 vehicles over ten months. BMW also describes the completed work as both a pilot and a successful deployment. That combined wording is useful. The program began as an evaluation, then crossed into operational production work.

The vehicle count should be read carefully. It does not mean the robot assembled 30,000 complete vehicles. Figure 02 performed one sheet-metal loading workflow that contributed to the production process for those vehicles. The direct robot output disclosed by Figure is the part count, not complete cars.

Targets are not measured outcomes

Figure says the team defined three critical key performance indicators:

1. Cycle time: 84 seconds total, including 37 seconds for loading.
2. Placement accuracy: more than 99% of cycles correctly loading all three parts per shift.
3. Interventions: zero human pauses or resets per shift.

These are valuable disclosures because they reveal what the program was trying to optimize. They are not the same as a public result table.

The source does not publish:

• median, lower-percentile, and worst-case cycle time;
• the number of shifts that met the 84-second requirement;
• attempted cycles and correctly completed cycles by shift;
• the achieved placement-success distribution;
• total human pauses, resets, or recoveries;
• intervention time or reason categories;
• excluded cycles, rejected parts, rework, or damaged parts.

Figure’s conclusion says every intervention was logged. That establishes the existence of an internal record, not its public value. A statement that the goal was zero interventions cannot be converted into a claim that every shift achieved zero.

This distinction matters for buyers. A robot can load 90,000 parts and still require an unknown amount of support. The deployment remains meaningful, but staffing, reliability, and economics cannot be reconstructed without the assistance record.

Runtime and schedule still need fleet denominators

More than 1,250 runtime hours is substantial for a public humanoid deployment claim. It gives readers a time measure instead of only a selected video or cumulative part count.

The humanoid operating-hours guide shows how to keep runtime, uptime, utilization, productive time, and fleet-summed robot-hours from being treated as synonyms.

The source does not publish the robot count behind that total. It uses both singular and plural language at different points, including a statement that robots were delivered to the plant. Without a fleet denominator, the public record cannot establish runtime per robot, simultaneous operation, or utilization.

The stated 10-hour weekday shift is also a schedule description, not a complete uptime result. A normalized record would separate:

• scheduled hours;
• available hours;
• active motion and productive hours;
• material-starved or line-blocked time;
• charging and planned service;
• unplanned downtime;
• intervention and reset time;
• completed output during each time class.

Figure says full deployment ran every working day. That supports recurring operation. It does not establish that the robot was active for every scheduled minute or that all scheduled shifts met the task targets.

What the deployment taught about hardware

Figure identifies the forearm as Figure 02’s leading hardware failure point at BMW. The subsystem combined tight packaging, three degrees of freedom, thermal constraints, a local circuit board, communications, and dynamic cabling.

For Figure 03, the company says it removed the forearm distribution board and dynamic cabling so each wrist motor controller communicates directly with the main computer. Figure presents this as a reduction in complexity intended to improve reliability and thermal management.

This is useful design-trace evidence. A field problem was identified, and the successor architecture changed in response. It is not yet a measured Figure 03 reliability gain. Stronger evidence would compare matched failure rates, operating hours, thermal events, maintenance, and repair time between generations under relevant loads.

The lesson is broader than one wrist. Production operation can reveal failure modes that a short laboratory evaluation misses. The value of those lessons becomes measurable only when the successor publishes its own field record.

Figure 03 changes both the robot and the task

BMW’s next project moves from body-shop sheet-metal loading to logistics sequencing in Hall 52. Components arrive unsorted in larger containers. Figure 03 is expected to pick them and sort them into a sequencing trolley, then move the trolley to a collection point. Another transport system takes the sequenced parts to assembly employees.

Figure’s demonstration adds several visible control demands:

• detect parts whose positions and orientations vary;
• grasp thin-walled components with both hands;
• step and reposition while maintaining reach and balance;
• place each part into the correct trolley slot;
• pull a large cart on caster wheels;
• alternate between precise placement and higher-force cart movement.

Figure attributes the behavior to Helix 02, its vision-language-action system for coordinating hands, arms, torso, and feet. The company says high-frequency visual-motor control adapts motion to spatial variation and corrects small errors during execution.

The demonstration establishes that Figure 03 performed the shown behavior in a BMW manufacturing environment. It does not establish how many attempts were made, whether the sequence was continuous or edited, what variants were tested, how often errors occurred, how human help was used, or whether the workflow met BMW’s production acceptance criteria.

BMW’s language is slightly different from Figure’s. BMW says it looks forward to deploying Figure 03 for the sequencing use case and describes what the robot will do. Figure calls the released evidence the first demonstration. Together, the sources establish customer-backed deployment intent plus a site demonstration. They do not yet establish recurring production operation.

Why the Figure 02 record cannot transfer to Figure 03

Figure 02 and Figure 03 share a company, customer, and site, but they differ in ways that affect evidence:

• hardware generation;
• wrist and forearm architecture;
• hands and sensing;
• charging approach;
• safety-oriented physical design;
• model and control stack;
• workflow;
• materials and container presentation;
• locomotion and cart interaction;
• task acceptance criteria.

Figure 02’s record proves that Figure and BMW brought one humanoid generation into one bounded production task. It also provides relevant integration experience, site knowledge, calibration methods, and failure data.

It does not prove that Figure 03 can match Figure 02’s runtime, weekday schedule, placement quality, intervention burden, maintenance, or safety performance. It especially does not prove reliability on sequencing, which combines variable part selection, two-handed handling, whole-body repositioning, and cart movement.

This generation boundary is also why our humanoid availability tracker keeps retired Figure 02 production evidence separate from current Figure 03 status.

Safety language needs an application record

BMW says the Figure 02 collaboration demonstrated that humanoid robots can safely perform precise, repeatable steps under real production conditions. It also describes Figure 03 soft components as designed to enhance safety.

Those are first-party safety statements. The reviewed sources do not publish:

• the application risk assessment;
• applicable standards and editions;
• protective functions and validated performance levels;
• speed, force, stopping, and separation test results;
• payload-dependent limits;
• safety stops, near misses, incidents, and exclusions;
• controls for cart motion, dropped parts, pinch points, and shared access;
• revalidation after hardware or software changes.

A customer statement that work occurred safely is stronger than a visual impression from a video. It is still not a complete safety case. The relevant unit is the whole application, including robot, part, trolley, fixtures, transport systems, people, layout, operating modes, and recovery procedures. Our robot safety standards guide explains how to map that application-level evidence.

Commercial and economic evidence remains private

The BMW program clearly represents a substantial customer relationship. Public sources establish a multi-generation collaboration and real production work. They do not disclose whether the Figure 02 work or Figure 03 phase uses a paid contract, lease, service agreement, development partnership, or another commercial structure.

The sources also omit:

• robot purchase or service price;
• integration and facility cost;
• vendor and BMW staffing;
• maintenance and spare inventory;
• energy and charging cost;
• line downtime caused or avoided;
• throughput and quality against the prior process;
• ergonomic or safety outcomes with denominators;
• service levels, guarantees, and renewal terms.

BMW presents humanoid robotics as a complement to existing automation, especially for monotonous, ergonomically demanding, or safety-critical activities. That is a strategic rationale, not a return-on-investment result.

A defensible economic comparison would use the same volume, quality, schedule, staffing, safety, and facility boundaries for Figure 03 and the process it replaces. Production capability and economic advantage are separate claims.

What a complete Figure 03 production record would contain

A useful next disclosure would cover one fixed reporting window after customer acceptance.

Evidence block	Minimum useful disclosure
Scope	Site, line, part families, container variation, trolley definition, collection point, and exclusions
Fleet	Robot count, hardware revision, end effectors, Helix version, and changes during the period
Time	Scheduled, available, active, productive, charging, maintenance, blocked, and down hours
Output	Part attempts, correct placements, wrong-slot events, dropped or damaged parts, completed trolleys, and retries
Timing	Full sequence, pick, placement, cart-move, and lower-percentile cycle times
Assistance	Teleoperation, interventions, resets, recoveries, remote support, and on-site staffing
Reliability	Faults, failure categories, mean time between failures, repair time, and replaced components
Safety	Stops, near misses, incidents, validated limits, and corrective actions
Quality	Customer acceptance criteria, audit method, rejected output, and traceability
Baseline	Prior manual or automated process with matched volume, quality, labor, and schedule
Economics	Integration, service, support, energy, maintenance, facility changes, and commercial terms
Verification	BMW confirmation and a method that reconciles source logs with published totals

This table would let readers distinguish a successful demonstration from reliable production. It would also make Figure 03 comparable with the bounded deployment cases in our physical AI production evidence guide.

Evidence boundaries

Claim	Classification	Confidence	Boundary
Figure 02 supported more than 30,000 BMW X3 vehicles	Customer-confirmed production context	High	The robot performed one sheet-metal loading step, not complete vehicle assembly
Figure 02 loaded more than 90,000 parts and ran for more than 1,250 hours	Vendor-reported operating result	High that Figure published the totals	Fleet size, attempt count, intervention, downtime, and audit are not public
Figure 02 met an 84-second cycle	Not established as an achieved distribution	High	Figure publishes the requirement, not achieved shift-level timing results
Figure 02 exceeded 99% placement success per shift	Not established as an achieved result	High	Figure publishes the target, not the measured distribution
Figure 02 completed every shift with zero interventions	Not established	High	Zero interventions was the goal; Figure says interventions were logged but does not publish them
Figure 03 performed the shown sequencing behavior at BMW	First-party site demonstration	High	Attempts, continuity, editing, assistance, variants, and acceptance thresholds are undisclosed
Figure 03 is already production deployed	Not established	High	BMW describes the next deployment phase, while Figure calls the evidence a first demonstration
Helix 02 is a proven general-purpose factory controller	Not established	High	One demonstrated sequencing workflow does not establish broad task generality or recurring reliability
Figure 03 is more reliable than Figure 02	Design intent, not measured field result	High	Architecture changes are disclosed, but matched failure and maintenance data are not

What to watch next

1. BMW confirmation that Figure 03 passed site acceptance and entered recurring logistics work.
2. Start date, reporting window, scheduled shifts, and robot count.
3. Part attempts, correct placements, completed trolleys, errors, retries, and damage.
4. Median and lower-percentile cycle times under real container variation.
5. Teleoperation, interventions, resets, recovery, remote support, and staffing.
6. Active hours, downtime, charging, maintenance, failures, and repair time.
7. Application-level safety validation, stops, near misses, incidents, and change control.
8. Matched results against the prior sequencing process.
9. Commercial terms, support obligations, and renewal or expansion.
10. Equivalent records for any additional Figure 03 workflows or sites.

Verdict

Figure 02 at BMW deserves a production-deployment classification. The record combines a named customer, active automotive line, bounded task, ten-month customer-confirmed duration, weekday operation, more than 1,250 runtime hours, and more than 90,000 parts. Few public humanoid cases combine that many evidence layers.

The record is still incomplete. Figure publishes task targets without the achieved cycle, placement, and intervention distributions. Fleet size, uptime, maintenance, safety denominators, staffing, and economics remain private.

Figure 03 is the next test, not the continuation of the same result. Its sequencing demonstration is technically broader than Figure 02’s pick-and-place workflow, and BMW publicly describes deployment intent. The new generation must now earn a new operating record.

The defensible conclusion is narrow and useful: Figure proved that a humanoid could perform one recurring production task at BMW. Figure 03 has demonstrated a new logistics task at the same site. Production status for the successor should wait for customer acceptance, repeated shifts, runtime, throughput, intervention, reliability, safety, and commercial evidence.

Frequently asked questions

Was Figure 02 deployed in BMW production?

Yes. BMW says Figure 02 supported production of more than 30,000 BMW X3 vehicles over ten months at Plant Spartanburg. Figure reports more than 90,000 parts loaded, more than 1,250 runtime hours, and operation on every working day during full deployment. This is production evidence for one bounded sheet-metal loading workflow, not proof of general factory autonomy.

Did Figure 02 achieve its cycle-time, placement, and intervention targets?

The public Figure source names an 84-second total cycle requirement, a target above 99% successful placement per shift, and a goal of zero human pauses or resets per shift. It does not publish achieved distributions or totals for those three measures. The runtime and throughput record is meaningful, but the target values should not be rewritten as measured results.

Is Figure 03 production deployed at BMW?

BMW describes Figure 03 as the successor that will start a logistics sequencing application, and Figure calls its June 2026 evidence the first demonstration of Figure 03 performing that workflow at Plant Spartanburg. The reviewed sources do not yet publish recurring Figure 03 shifts, runtime, throughput, intervention, uptime, or acceptance results.

What task does Figure 03 perform at BMW?

The planned sequencing workflow starts with components arriving unsorted in larger containers. Figure 03 picks and sorts them into a sequencing trolley, then moves the trolley to a collection point for onward transport to assembly. Figure also demonstrates handling thin-walled parts, stepping and repositioning during manipulation, and pulling a cart.

Can Figure 02's results be used as Figure 03 reliability evidence?

No. Figure 02's deployment gives Figure and BMW relevant operating experience, but Figure 03 is a different hardware generation performing a different workflow. The successor needs its own customer acceptance, runtime, task-success, intervention, maintenance, safety, and economics record.

What evidence should BMW and Figure publish next?

The most useful next disclosure would identify the Figure 03 fleet, fixed reporting window, scheduled and active hours, part attempts and correct placements, cycle-time distribution, interventions, resets, downtime, maintenance, safety events, staffing, baseline process, and customer acceptance result.