In the summer of 2024, Figure AI shipped a single humanoid robot — the Figure 02 — to BMW's sprawling Spartanburg factory in South Carolina, the automaker's largest plant worldwide, where X3, X5, and X7 models roll off the line. Eleven months later, that robot had logged 1,250 operational hours, moved over 90,000 sheet metal components across the welding floor, walked roughly 1.2 million steps through the factory, and directly contributed to the production of more than 30,000 BMW X3 vehicles.
Then it was retired.
The Figure 02 experiment at Spartanburg is one of the most data-rich, real-world deployments of a general-purpose humanoid robot in industrial history. And the honest assessment of what worked — and what didn't — is precisely the kind of evidence that North American VCs have been starved for as they place billion-dollar bets on a still-unproven category.
The Numbers That Matter
Let's start with the hard data. Over its 11-month deployment at BMW Spartanburg:
- 90,000+ components carried — sheet metal parts removed and positioned for welding operations
- ~1,250 operational hours on 10-hour shifts, 5 days per week
- 30,000+ BMW X3 vehicles touched by the robot's work
- ~1.2 million steps walked across the factory floor
These are not simulation numbers. They are not lab benchmarks. They are production-line telemetry from one of the world's most demanding automotive OEMs.
The robot performed a specific but tactically meaningful task: removing and positioning sheet metal parts within the welding process. It was not a replacement for a full production line, nor did it operate 24/7 unattended. But within its defined scope, it demonstrated that a bipedal, two-armed humanoid could integrate into an active automotive production environment and deliver measurable throughput.
"From lab to line transition happened faster than we expected," the company noted — a surprisingly candid acknowledgment from a startup racing against Tesla and others toward commercial viability.
What Worked, What Didn't
What worked: The Figure 02 proved that a humanoid robot can physically navigate a real automotive factory — stairs, narrow aisles, dynamic environments — and execute precision manipulation tasks without facility redesign. The robot integrated with BMW's existing MES (Manufacturing Execution Systems) and production workflows, requiring no new infrastructure beyond additional safety barriers and 5G connectivity upgrades.
What didn't work: Safety protocols had to be reinforced. Additional barriers were installed beyond initial expectations. The 5G connection proved critical — a reminder that humanoid robots are only as reliable as the network infrastructure they depend on. More importantly, the robot did not achieve fully autonomous 24/7 operation. It required supervision, intervention, and progressive learning cycles.
The retirement of Figure 02 is not a failure — it is a deliberate hardware generation cycle. Figure AI designed the 02 as a data-collection platform for its next-generation architecture. Every failure mode, every edge case, and every operational nuance encountered at Spartanburg was fed into the training pipeline for Figure 03 and the Helix neural architecture.
The BMW Dual-Track Strategy
Here is where the story gets strategically interesting — and where most coverage has missed the signal.
BMW is not putting all its eggs in the Figure basket. The automaker is pursuing what can only be described as a dual-track industrial humanoid strategy:
- The America Track (Figure AI): Spartanburg deployment completed → now evaluating Figure 03 for Leipzig factory deployment in Germany
- The Europe Track (Hexagon Robotics AEON): Signed separately for Leipzig factory as well, targeting high-voltage battery assembly starting December 2025
Two robots, two architectures, one factory. BMW is running a bake-off.
The strategic logic is clear: by 2026, BMW expects to have hard comparative data on two different humanoid platforms performing distinct tasks in the same production environment. One American (Figure 03 with Helix), one European (Hexagon AEON). The winner earns the inside track to scale across BMW's 30+ production facilities worldwide.
BMW has also established a formal "Center of Competence for Physical AI in Production" — a dedicated organizational unit tasked with evaluating and deploying embodied AI across its manufacturing footprint. This is not a skunkworks project. It is a structured, resourced initiative with executive sponsorship.
The timing is no coincidence either. BMW's incoming CEO, Milan Nedeljković, built his career in production. As the former head of production, he was the internal champion of the iFACTORY strategy — the company's three-pillar approach to lean, green, and digital manufacturing. Physical AI is the logical extension of the "digital" pillar. Nedeljković's elevation to the CEO role signals that BMW intends to double down on manufacturing innovation, with humanoid robots as a centerpiece.
Figure 03 + Helix: The Next Chapter
Figure 02's retirement made room for what Figure AI considers its true production platform: Figure 03, powered by the Helix neural architecture.
Helix itself represents a significant architectural departure. It splits control into two layers:
- "Brain" — A 7-billion-parameter Vision-Language Model (VLM) for high-level reasoning, scene understanding, and task planning
- "Cerebellum" — An 80-million-parameter Transformer running at 200Hz for low-level motor control, reflexive responses, and precision manipulation
This separation mirrors biological motor control — deliberative reasoning in the cortex, reflexive execution in the cerebellum. The "brain" sees and plans; the "cerebellum" executes with speed and precision.
The early results from Figure 03's commercial deployments suggest the architecture is delivering:
Catalyst Brands (JCPenney parent) warehouse logistics:
- 3 Figure 03 units operated for 200 continuous hours
- Processed ~250,000 packages with zero mechanical failures
- Each unit passed 80 functional validation tests before deployment
BotQ factory production metrics:
- 1 unit/hour production throughput
- 80% first-pass yield rate
- Target: ~12,000 units per year at scale
Figure AI's valuation has surged to $39B on the back of these milestones — placing it among the most valuable private AI companies globally. But the key metric for VCs is not valuation; it is whether Figure 03 can achieve 24/7 autonomous operation in BMW's Leipzig factory, which remains the next critical checkpoint.
The Industrial Benchmark
How does Figure's progress compare to the competitive landscape?
| Company | Industrial Deployments | Units Deployed | Key Metric |
|---|---|---|---|
| Figure AI | BMW (completed), Catalyst Brands (active) | ~dozens | 90,000 parts moved in real factory |
| Tesla Optimus | Tesla factories only (~8,000 units) | ~8,000 | Commercial maturity: 45.1/100 |
| 1X Technologies | Hayward, CA factory | Scaling to 10K/yr | OpenAI-backed; logistics focus |
| Boston Dynamics (Atlas) | Hyundai factory | Planning 30K/yr | Electric Atlas deployed in production |
The key insight: deployment count ≠ industrial validation. Tesla has deployed thousands of Optimus units, but exclusively in its own factories — which is very different from passing a demanding external OEM's validation process. Figure AI's dozen-ish units may have more external production-line data than any competitor.
1X Technologies is scaling its manufacturing facility in Hayward, California with a target of 10,000 units in its first year, backed by OpenAI. Boston Dynamics' electric Atlas has been deployed in Hyundai production lines, with plans to scale to 30,000 units per year.
Goldman Sachs projects the global humanoid robot market will reach $38B by 2035, with over 1.4 million units shipped annually. But those projections depend entirely on whether the current generation of humanoids can bridge the gap from supervised batch operations to fully autonomous continuous production.
Why This Matters for Investors
For North American VCs evaluating the humanoid robotics space, the Figure-BMW case study offers several actionable insights:
1. The "Factory as Laboratory" model works — but has limits. Real production data is the most scarce and valuable asset in embodied AI. Figure's Spartanburg deployment generated more real-world training data than any lab simulation could. But the costs — in engineering time, safety infrastructure, and deployment complexity — are significant.
2. 24/7 autonomy is the true north metric. Figure 02 did not achieve it. Figure 03 has not yet demonstrated it at BMW. The entire industry's valuation thesis rests on the assumption that humanoids can eventually operate untended in industrial environments. Every month that passes without this milestone is a month that supports the thesis that it may take longer than expected.
3. BMW's dual-track strategy reveals OEM procurement logic. Major automotive OEMs are running parallel evaluations — not because they lack conviction, but because they lack data. The winning humanoid platform will be the one that produces the best cost-per-operation numbers on an actual production line. BMW's Center of Competence for Physical AI is designed to generate exactly this kind of comparative intelligence.
4. The Leipzig deployment is the critical path. If Figure 03 achieves sustained, autonomous 24/7 operation at BMW Leipzig — especially alongside Hexagon's AEON platform in a comparative setting — it will be the most significant validation event in the industrial humanoid space to date. If it falls short, the $39B valuation will face renewed scrutiny.
5. Figure's architecture bet (Helix's brain-cerebellum split) is the right one, but unproven at scale. The VLM + high-frequency motor controller architecture is logically sound and aligned with neuroscience principles. But translating 200 continuous hours at Catalyst Brands into months of untended autonomous operation at BMW is a fundamentally different challenge.
The Bottom Line
The Figure 02's 1.2 million steps through BMW Spartanburg were not just a proof of concept. They were the first chapter of a dataset that will define whether humanoid robots graduate from supervised novelty to industrial necessity.
The robot is retired. But the data it generated — 90,000 components, 30,000 vehicles, 1,250 hours of production-line telemetry — is now training the next generation. And at BMW Leipzig, the real test begins.
