Figure 03 Hits 1 Robot Per Hour: Inside Figure AI's Mass Production Sprint
The headline: Figure AI’s BotQ manufacturing facility in California has gone from producing 1 Figure 03 per day to 1 per hour — a 24x throughput jump in under 120 days, with 350+ third-generation humanoids delivered to date. That single metric reframes the entire humanoid race. Figure isn’t just ahead on AI (Helix) or deployment (BMW); it’s ahead on the boring, hard, expensive thing every other humanoid company is going to need next: factory output. This post walks through the BotQ ramp, the BMW Spartanburg-to-Leipzig expansion, the Figure 02 retirement, and the new Helix System 0 stair-traversal breakthrough.
🔗 See the robot: figure.ai — Figure AI’s official site, with current Figure 03 demos, the BotQ production updates, and the Helix AI announcements.
The 24x manufacturing leap
Most humanoid companies in 2026 are still hand-assembling units in machine shops. Boston Dynamics is producing Atlas at “Boston HQ” with all 2026 units pre-committed before a single one ships. Tesla just announced production starts this summer at Fremont. 1X opened a Hayward, CA factory at the start of May.
Figure quietly did something different. In a recent update, the company disclosed that BotQ — its purpose-built humanoid manufacturing facility — has transformed from a prototype line into a high-output factory. The numbers from Figure’s own writeup:
- 24x throughput improvement in under 120 days
- Over 350 Figure 03 robots delivered from the line
- End-of-line first-pass yield exceeding 80%
- Battery production yield at 99.3% with 500+ units shipped
- 9,000+ actuators produced
- 150+ networked workstations running custom manufacturing software
- 80+ functional tests per robot including stress squats, shoulder presses, and jogging burn-in
The numbers matter for one reason that has nothing to do with bragging rights. Each robot rolling off the BotQ line is a data-collection engine. Figure is generating real-world manipulation data — the kind that can only come from physical robots operating in physical environments — at a rate that’s now over an order of magnitude faster than competitors. That data trains Helix, Figure’s vision-language-action model, which makes every subsequent robot smarter.
BotQ’s design capacity is 12,000 units per year, scaling to 50,000 in 2026’s later half and a long-run target of 100,000 per year. For comparison: Hyundai’s planned 30,000-units/year Atlas factory doesn’t come online until 2028. Tesla’s Fremont conversion is targeting low-volume Optimus production starting late July or August 2026. Figure has the head start, and it’s compounding.
BMW Leipzig is the next deployment
In February 2026, BMW announced that humanoid robots are coming to a second BMW plant — this time in Leipzig, Germany. The pilot follows the Spartanburg, SC deployment that ran from late 2024 through 2025 and which became the most-cited proof point in commercial humanoid robotics.
The Spartanburg numbers from BMW’s own press releases:
- 30,000+ BMW X3 vehicles supported in production
- 90,000+ sheet metal parts handled
- ~1,250 operating hours across roughly 10 months
- ~1.2 million steps logged
- 10-hour shifts, 5 days a week for the duration of the pilot
This isn’t BMW vendor-marketing the deployment. These are BMW’s own published metrics, which means they survived a corporate communications team that doesn’t have an incentive to overstate. That’s a level of validation no other humanoid company has come close to.
The Leipzig deployment, according to BMW’s announcement, ran initial laboratory tests in December 2025, then a deeper integration test in April 2026, with a full pilot phase scheduled for summer 2026. The deployment in Leipzig targets a multifunctional application — high-voltage battery assembly and component manufacturing — which is more complex than Spartanburg’s sheet-metal kitting. Per BMW’s own language, the Leipzig pilot is meant to test “wider hand and gripper attachments and dynamic use on wheels.”
The customer-initiated geographic expansion is the part that matters. BMW didn’t have to do this. They chose to. In enterprise hardware, customer expansion is the hardest signal to fake.
Figure 02 retired after a successful run
Alongside the Figure 03 production ramp, Figure announced in November 2025 that Figure 02 — the generation that did all the BMW Spartanburg work — is being officially retired. Units are returning to Figure HQ from BMW as part of a fleet-wide retirement.
The Figure 02 retirement post is worth reading for the hardware lessons it surfaced. The most important: Figure 02’s forearm was the top hardware failure point at BMW. The tight packaging, three-degree-of-freedom dexterity requirements, and thermal constraints combined into a subsystem that simply wore out faster than the rest of the robot. For Figure 03, the entire wrist electronics were re-architected — each wrist’s motor controller now communicates directly with the main computer, eliminating both the distribution PCB and the dynamic cabling that caused failures.
That kind of post-mortem only happens when you’ve run real shifts. It’s a level of operational maturity that’s rare in this industry.
Figure 03 itself was announced in October 2025 and incorporates the full Figure 02 deployment learnings. Key changes that came out of the BMW data:
- Soft textile covering — washable, replaceable, safe for home contact (yes, Figure is targeting homes too, not just factories)
- Wireless charging — Figure 03 steps onto a charging pad instead of plugging in
- Palm cameras — embedded in each hand for close-range visual feedback
- 3-gram tactile sensitivity — can feel a paperclip’s weight
- Re-architected wrists — eliminated the failure-prone PCB and cabling
- 2x camera frame rate, 75% lower latency, 60% wider FOV on the sensor suite
- Target consumer price ~$20,000 at full BotQ scale — comparable to 1X Neo and Tesla Optimus’s targets
Helix System 0 — zero-shot stair traversal
The other recent Figure announcement is on the AI side. The Helix System 0 (S0) update demonstrates perception-conditioned whole-body control trained entirely in simulation, transferred zero-shot to the physical robot. No real-world fine-tuning, no domain-specific calibration, no operator-in-the-loop adjustments.
The first demonstrated capability: stair traversal. The same neural network weights that learned to climb procedurally generated staircases in simulation now traverse real stairs on Figure 03 with the same human-like gait and recovery characteristics. Varying lighting conditions, different stair geometries — the policy generalizes.
Why this is significant: the sim-to-real gap has historically been the bottleneck for perception-driven robot control. Every other humanoid program does some amount of real-world fine-tuning to bridge sim-to-real. If Figure has genuinely eliminated that gap for whole-body control, the development pace for new behaviors should accelerate dramatically. Stair climbing is the demo; the architecture is what matters.
This connects to our deeper coverage of Figure’s Helix-02 multi-robot bedroom-reset demo, where the AI architecture is broken down in more detail.
What this all adds up to
A useful way to think about Figure right now: it’s the only Western humanoid program with all three pieces aligned simultaneously.
- AI in-house and at the frontier — Helix is end-to-end, RL-trained in simulation, transfers zero-shot. No other major humanoid program has demonstrated comparable sim-to-real generalization.
- Real-world commercial deployment with published metrics — BMW Spartanburg’s 30,000-X3, 90,000-part, 1,250-hour numbers come from the customer, not the vendor. Leipzig is the geographic expansion.
- Mass-manufacturing infrastructure operating now — BotQ at 1 robot per hour, 350 units delivered, designed for 12K→50K→100K annual capacity, all before the next major competitor even starts volume production.
The other thing worth flagging is the financial position. Figure raised a Series C in September 2025 at a $39 billion valuation, with NVIDIA, Microsoft, Jeff Bezos, and OpenAI all participating. Total raised: over $1.9 billion. That’s the kind of capital pile that lets you build BotQ before you have customers, lets you eat the per-unit losses while the cost curve comes down, and lets you outlast the slower competitors.
For our broader comparative take on where Figure 03 sits among the top humanoid robots in 2026, see our Top 5 Humanoid Robots of 2026 roundup. And for the AI architecture story specifically, the Helix-02 deep dive walks through how two Figure robots coordinate without a shared planner.
Bottom line
Figure 03 isn’t winning because it’s the most agile humanoid (Atlas) or the cheapest (Unitree) or the first in someone’s home (Neo). It’s winning because it’s the only humanoid program where the AI, the deployment, and the manufacturing are all running at production scale simultaneously. The other companies will catch up on individual axes. Catching up on all three at once is a much harder ask.
The next 12 months will reveal whether Figure can maintain this lead or whether Tesla’s vertical integration eventually closes the gap. For right now, BotQ at 1 robot per hour is the most important fact in commercial humanoid robotics.
