The most consequential thing that happened in robotics in the first half of 2026 was not the demonstration of a new capability. It was a demonstration that an old capability had become routine. On May 13, Figure put its humanoid robots on a package-sorting conveyor belt and let them work an eight-hour shift, processing barcoded packages at speeds the company says match the throughput of a human operator. Eight hours of continuous, autonomous, productive work, with no resets, no intervention, no choreography. The robots clocked in, did the work, and clocked out. That is not a research milestone. That is a labor milestone, which is a different category of event, and the implications run in different directions.
The capability behind this is a system Figure calls Helix-02, which is the second generation of a vision-language-action model the company has been building since early 2025. The first Helix was an upper-body system — control of arms, hands, and the muscles that handle the dexterity-heavy work of picking, placing, and manipulating objects. Helix-02 extended that control to the entire robot. Legs, torso, arms, fingers, head, and the coordination between them, all unified under a single model rather than handed off between specialized subsystems. The architecture choice is not glamorous, but it is the choice that makes eight-hour autonomy possible. A robot whose upper body is controlled by one model and whose legs are controlled by another has two systems that can disagree, two systems that can fail independently, and two failure modes that need to be handled at the boundary. A robot whose entire body is controlled by one unified model has one system, one set of failure modes, and one coherent representation of what the robot is currently doing and why.
The thing that makes Helix-02 work, which the company has been increasingly willing to talk about, is a foundation layer they call System 0. System 0 replaced what the engineering team described as more than a hundred and nine thousand lines of hand-engineered C++ locomotion code with a neural controller trained on more than a thousand hours of human motion data. That replacement, in itself, is the kind of change that would have been considered reckless eighteen months ago and is now considered table stakes for any serious humanoid robotics program. Hand-written locomotion code is brittle. It works under the conditions the engineers anticipated and fails in unexpected ways when the conditions shift. A neural controller trained on enough variety of human motion is, perhaps counterintuitively, more reliable, because it has seen more of the long tail of conditions a real environment produces and has learned to handle them as instances of patterns it already understands rather than as exceptional cases that need exceptional handling. The transition from hand-engineered control to learned control is the same transition that happened in computer vision a decade ago, then in natural language processing, and is now happening in robotics on a longer timeline because the data is harder to collect.
The data problem is the thing that has historically separated robotics from the other domains where learned systems have taken over. In language modeling, the data was already there, in the form of the web. In computer vision, the data was already there, in the form of every photograph ever uploaded. In robotics, the data does not exist by default. Every hour of robot training data has to be collected, by a robot, doing the thing the robot is going to be asked to do, in the kind of environment the robot is going to be asked to do it in. That is the fundamental reason robotics moved slower than the adjacent fields for so long. It is also the thing Figure has been quietly fixing. The company opened a new facility this spring called the Helix Lab, dedicated specifically to large-scale data collection for the Helix model. The lab captures egocentric human video — first-person footage of people doing the kinds of tasks the robots will eventually do — along with interaction data from the robots themselves working in varied environments. The output of that operation is the training fuel that makes the next Helix generation possible, and the rate at which the lab can produce that fuel determines the rate at which the model can improve.
This is the architectural pattern that makes Figure look more like a software company than like a hardware company, and it is the pattern worth dwelling on because it explains a lot of what is happening at the field's leading edge. A traditional robotics company would have a hardware engineering team, a controls team, a perception team, and a manufacturing operation. A software-shaped robotics company has a hardware engineering team, a foundation models team, a data operations team, a manufacturing operation, and a product team that sits on top and turns model improvements into customer-visible capability. The center of gravity shifts. The hard work moves from designing each individual subsystem perfectly to building the data pipeline that produces a model that handles all the subsystems together. The number of moving parts goes down. The number of engineers required to make the robot work goes down with it, even as the robot's capability goes up. That is the same productivity curve software companies have been riding for two decades, and it is now arriving in physical machines.
The manufacturing side of the story has been moving at a pace that does not get the attention it deserves, because manufacturing news is structurally less exciting than robotics demonstrations. Figure's third-generation humanoid, the Figure 03, started serious production this spring at a facility called BotQ. In under a hundred and twenty days, the company moved from a production rate of one Figure 03 per day to one Figure 03 per hour — a twenty-four-fold throughput improvement in less than four months. They have delivered more than three hundred and fifty units of the third generation. That throughput number is the one to watch, because it is the number that tells you whether the robots are an interesting demo or an emerging product line. Three hundred and fifty units in a few months is the kind of volume that crosses the line from research to commercial. It is small compared to a car factory, but it is large compared to anything anyone was producing in humanoid robotics two years ago, and the slope of the curve is what matters more than the absolute number. If the rate keeps doubling every few months — which is what the BotQ trajectory implies — Figure will be producing thousands of units a month by the end of 2026 and ten thousand units a month within another twelve months after that.
The other demonstration the company ran this spring is the one worth describing carefully because it captures what changes when a humanoid can actually do continuous work. A Figure robot walked from a starting position to a dishwasher, opened the dishwasher, unloaded the dishes, navigated across a kitchen, stacked the items in cabinets, returned to the dishwasher, reloaded it with dirty items, and started the cycle. End to end, no human intervention, no resets, no choreography. The whole sequence took about four minutes. Figure described it as the longest-horizon autonomous task ever completed by a humanoid, and the description is technically accurate but undersells the change. The thing that matters about this demonstration is not that any individual step was hard. Walking is solved. Manipulating dishes is largely solved. Stacking items is mostly solved. What is hard is sequencing the steps reliably, recovering from the small errors that occur in every step, and maintaining coherent intent across a multi-minute task. Four minutes of coherent intent in a humanoid is the threshold at which the robot starts being useful for real work, because real work consists of multi-minute tasks that need to be performed reliably enough that a human does not need to supervise. Eight hours of package sorting is a stress test of the same capability — it is the same skill, scaled up to a full shift, and the fact that the robot did not break down or wander off or require intervention is the evidence that the capability is robust enough to deploy.
The economics of all this depend on a number that Figure has not yet disclosed but that determines almost everything: the marginal cost of producing one additional Figure 03, and the operating cost of running one for a full year of shifts. Public estimates put the manufacturing cost in a range that, if it can be reduced through scale, would put the total cost of ownership for a humanoid robot somewhere below the annual cost of a human warehouse worker in a developed country within the next twenty-four months. That is the crossover line. When the robots become cheaper than the humans they replace for a meaningful range of tasks, the deployment dynamics change qualitatively. The current pilot installations — packaging, warehouse work, light manufacturing assist — are early commercial deployments, not just research demonstrations. The customers running them are looking at the same numbers, doing the same math, and the answers they get back will determine how aggressively they expand the deployments. If the math works at the current production cost, the deployments expand to the next set of customers as soon as Figure can ship units. If it does not yet work, the deployments stay narrow until the unit cost comes down. Either way, the rate of cost reduction over the next twelve months is the most important number in the company's future, and the BotQ throughput improvements suggest the team is taking the cost question seriously enough to be making real progress on it.
There is a question that comes up in every conversation about humanoid robotics this year, which is whether the right form factor is even humanoid, or whether the field is letting aesthetic preference override engineering pragmatism. Wheels are more efficient than legs for most warehouse work. Specialized arms are more dexterous than generalized hands for most manipulation tasks. A purpose-built picking robot for one specific task will outperform a humanoid trying to do the same task at the same price point. The argument for humanoids, which Figure has been making consistently, is that the world is built for humans. Doorways, stairs, vehicles, tools, workspaces, the height of shelves, the orientation of controls — all of it assumes a human body. A humanoid robot can move into that existing environment without modification, do the work the environment was designed for, and use the tools that humans already use. A non-humanoid robot needs the environment redesigned around it, which is expensive and slow and often impractical for retrofitting existing facilities. The humanoid form factor is the universal interface, and once the cost of producing it drops below the cost of redesigning the world, it becomes the rational default.
That argument is becoming harder to dismiss as the units start working full shifts. A humanoid that can do an eight-hour package sorting shift today can plausibly do an eight-hour shift at a different task tomorrow, and the cost of moving it between tasks is the cost of retraining the model on new data, not the cost of redesigning the hardware. That is the leverage point. One hardware platform, one foundation model, an infinite range of jobs as the data accumulates. The Helix Lab is the bet that the data will accumulate fast enough to justify the platform investment. The BotQ ramp is the bet that the hardware will be cheap enough at scale to make the units economically interesting. The May demonstrations are the evidence that both bets are paying off faster than the field's median analyst expected.
The thing to watch over the next twelve months is whether Figure can keep the production rate climbing while keeping the unit cost falling, whether the Helix Lab can generate enough training data to maintain the rate of capability improvement that Helix-02 represented over Helix-01, and whether the first wave of commercial deployments produces customer stories strong enough to pull the second wave in faster. None of those are guaranteed. All three are plausible on the current trajectory. The interesting variable, more than any of them individually, is the integration: a hardware company with a strong manufacturing operation and a software company with a strong training pipeline have historically been different kinds of companies, and Figure is trying to be both at once. If the integration works, the result is a category of company that does not really exist yet — a robotics company built like a software lab, with the manufacturing scale of a vehicle maker and the capability improvement cadence of a model lab. That combination is what the field has been waiting for since the first wave of humanoid demonstrations a decade ago, and May 2026 is when it stopped being theoretical.
Robotics companies running fleets of learned-control humanoids share the same orchestration problems as any large-scale model deployment: routing, observability, version management, cost attribution. Explore Swfte Connect for the infrastructure layer that lives across both worlds. For the broader May 2026 model landscape, see our AI API pricing trends post.
Sources: