For most of the last decade, the binding constraint in humanoid robotics was the demonstration. Could the machine fold a shirt, sort a bin, survive a trade show floor without a handler? Those questions are mostly answered. In 2026 the binding constraint has moved upstream, into the least glamorous part of the industry: the factory. The companies that will define the next five years are being selected right now by cycle times, first-pass yields, and actuator SKU counts, and this week Figure offered the most detailed look yet at what that selection looks like from the inside.
In a manufacturing update posted to its newsroom, Figure disclosed that its high-volume facility, BotQ, has delivered more than 350 third-generation humanoid robots and increased its production rate from one Figure 03 per day to one per hour. That is a 24x throughput improvement in under 120 days. The company frames this as a development-velocity story rather than a manufacturing story, and it is right to: every unit that leaves the line is simultaneously inventory, a data-collection engine, and a node in a fleet operations stack. But the raw production numbers deserve scrutiny on their own terms, because they are the first public, quantified account of what “scaling a humanoid factory” actually means.
The anatomy of a 24x ramp
The headline number, one robot per hour, is the least interesting figure in the update. The interesting ones describe everything that had to be true for the line to hit that cadence.
BotQ now runs dedicated production lines for all critical modules of the robot, orchestrated by custom manufacturing execution software across more than 150 networked workstations. Incoming quality is enforced by qualifying hundreds of suppliers against inspection criteria before parts enter the building. More than 50 in-process inspection points sit along the build process, and every finished robot must pass more than 80 functional verification tests before sign-off, including multi-limb stress testing and burn-in sessions in which robots perform squats, shoulder presses, and jogging at cycle counts in the thousands to shake out early-life failures.
The yield figures are the real reveal. Figure reports an end-of-line first-pass yield above 80% and improving weekly, a battery line at 99.3% first-pass yield with more than 500 battery packs shipped, and more than 9,000 actuators produced across more than 10 distinct SKUs.
Pause on that actuator number. A fleet of roughly 350 robots at approximately 25 to 30 actuated degrees of freedom per machine accounts for perhaps 9,000 to 10,500 actuators installed. In other words, actuator production volume and robot delivery volume are almost exactly in balance, with little visible surplus for spares yet. Actuators, not robots, are the true volume product of a humanoid factory, and the supply chain, inspection infrastructure, and rework loops built around them are where the cost curve will be won or lost. An 80% end-of-line first-pass yield sounds respectable for a young product, but it also means one in five robots currently requires rework before it can ship, and rework is where labor hours and margin quietly evaporate. The company’s own framing, “improving weekly,” is an admission that the line is still climbing the learning curve, not at the top of it.
Robots as a fleet operations problem
The second half of Figure’s update is about what happens after robots ship, and it reads less like a press release than like the field-service playbook of an equipment OEM. Robots rolling off BotQ are allocated across internal R&D, data collection, end-to-end housework development, and commercial use cases. Running more robots for longer exposed failures invisible at smaller scale, and the company responds with the standard instruments of mature fleet management: a diagnostics and failure-analysis system that pins root causes in minutes, software “fallback ladders” that let a robot gracefully degrade rather than halt on non-critical faults, an internal field service management system, fleet-wide OTA update infrastructure, and formal processes for upgrades and recall campaigns.
This is the unglamorous layer that separates a robotics program from a robotics business. The industry’s historical failure mode was not bad demos; it was fleets that could not be kept alive. Anyone who ran warehouse AMR programs in the 2010s knows the pattern: a pilot of 20 robots behaves beautifully, a fleet of 200 drowns the service organization. Figure’s explicit focus on the “long tail” of edge-case failures, a stage it says “only comes with significant fleet hours,” is the correct read of where the industry actually is. Capability per robot is rising fast, but the durable moat is accumulating operating know-how per fleet-hour, and that asset compounds only with physical scale.
There is also a subtle financial claim buried in the operations section. If robots ship to customer sites and residential homes while field failures feed hardware revisions, then Figure is running a closed-loop hardware iteration cycle measured in weeks, not model years. That is the same loop that made automotive and consumer electronics incumbents nearly impossible to displace, and it is precisely the loop that a two-year hardware generation cycle, the one Figure itself just demonstrated by melting its F.02 fleet, puts under pressure. The factory strategy and the depreciation treadmill are two ends of the same machine: faster ramps make hardware cheaper per unit but also shorten the economic life of every unit already in the field.
S0 learns to see: the quiet sim-to-real milestone
Tucked below the manufacturing detail is the update’s most technically significant news, and it has nothing to do with the factory. Figure’s System 0, the whole-body control layer of its Helix stack, has gained perception conditioning.
Until now, S0 reasoned only about the robot’s own body: joint states, base motion, proprioception. It walked flat ground confidently but was effectively blind. Stairs, ramps, and uneven terrain required hand-tuned mode switches and operator intervention. The fix is architecturally clean: RGB images from the head cameras pass through Figure’s stereo model, which lifts them into a 3D representation of the scene, and that spatial understanding is fed to the locomotion policy alongside proprioceptive state. The policy is trained end-to-end with reinforcement learning in simulation across thousands of randomized terrains, and the same network weights that climb procedurally generated staircases in simulation traverse real stairs on hardware. The transfer is zero-shot: no real-world fine-tuning, no domain-specific calibration, no operator in the loop, across varying lighting conditions.
Figure’s claim that “the sim-to-real gap is no longer the bottleneck for this class of behavior” should be read precisely. This is not a claim that sim-to-real is solved for manipulation, nor for long-horizon autonomy. It is a claim about one specific class: loco-manipulation policies where the scene geometry matters, trained in sim and deployed without adaptation. Stairs are the demonstration, but the underlying capability is general, and stairs are a hard requirement for the home. It also lines up with the company’s September release of Helix 2.5, which Figure says delivered zero-shot generalization of whole-body household behaviors across roughly 30 test homes. Stack the two together and the trajectory is clear: perception-conditioned lower-body control plus language-conditioned task policies, converging toward a single generalist stack that neither mode-switches nor teleoperates.
The money is chasing exactly this
The timing of Figure’s disclosure is not accidental. Capital is pouring into the sector at a pace that makes even 2025 look conservative. Dealroom’s public humanoid tracker puts VC investment into humanoid robotics startups at $11 billion in 2026 year-to-date, more than double 2025’s full-year record, with the largest rounds concentrated in Neura Robotics, XPeng Robotics, AI2 Robotics, Sharpa, and Apptronik.
The most striking of these is Germany’s NEURA Robotics, which raised a round of up to $1.4 billion with backing that now includes Amazon, Nvidia, Tether, Qualcomm, Bosch, Schaeffler, and the European Investment Bank, with the capital earmarked explicitly for mass production and its Neuraverse software ecosystem. Note the composition: strategic buyers from automotive supply chains, a stablecoin issuer, and a public lending institution. That is not a venture bet on a demo; it is an industrial-policy-adjacent bet on manufacturing capacity. The UK’s Humanoid, meanwhile, closed a $152 million Series A at a $1.35 billion valuation, a reminder that the financing window now opens for pretenders as readily as for leaders, which is itself a late-cycle signal.
On the volume front, the Chinese fleet numbers continue to dwarf Western deployments. Industry tracker Humanoid Press reports an AgiBot volume surge past 20,000 units with an IDC revision to roughly 25,000 for the first half, an advanced mass-production target for XPeng’s IRON, and UBTech U1 consumer deliveries underway. In the consumer segment, 1X says its October 2025 NEO launch booked a full year of production capacity, some 10,000 robots, in five days, with US deliveries beginning in 2026 and an EQT partnership covering another 10,000 units. And Tesla continues to promise Optimus V3 production in Fremont “soon,” with guidance tempered on its latest earnings call.
Read together, these are not separate stories. They are one race with three lanes: Western capital intensity (Figure, Apptronik, NEURA), Chinese manufacturing volume (AgiBot, XPeng, UBTech), and consumer-facing early revenue (1X, UBTech). Figure’s BotQ update is best understood as the first Western company publishing hard manufacturing data in an attempt to compete in the lane where the Chinese players are currently strongest.
What to watch next
First, yield disclosures will become the industry’s new due-diligence metric. Anyone can quote a run rate for a week; first-pass yield, inspection point counts, and actuator SKU volumes cannot be faked for long. Expect serious investors to demand exactly the figures Figure just volunteered.
Second, the IEEE-RAS International Conference on Humanoid Robots, October 16-18 in Santa Clara, lands in eight days with the sector at peak funding. Watch for perception-conditioned whole-body control papers to multiply now that Figure has shown the architecture works zero-shot on hardware.
Third, the tension between hourly production and two-year hardware lives will force the depreciation question into the open. Figure’s own numbers now contain both halves of the problem: a factory that can refresh the fleet annually and a business model that must recover robot cost over a shrinking service life. The company that resolves that tension, through durability engineering, resale channels, or software subscriptions that outlive the metal, will own the category’s economics.
The demo era is over. The manufacturing war has begun, and the first battle reports are yield tables.