Humanoid Robots in 2026: The Production Line, the Pilot, and the Press Release

7 min read

Key Claim: In 2026, humanoid robotics has one production benchmark — Figure’s BMW trial — and a substantial gap between that verified standard and the claims made by most other companies in the sector.

Two robots. One task. 99% accuracy. Eleven months. That is the clearest data point the humanoid robotics industry has produced so far, and it came from a sheet-metal loading trial at BMW’s Spartanburg plant — not from any of the market projections, funding announcements, or stage demonstrations that have dominated coverage of the sector. Understanding where that bar sits, and which companies are above or below it, is the most useful thing an engineering or procurement team can do before the vendor calls start.

The humanoid robotics sector attracted over $6 billion in investment in 2025. Global shipments totalled approximately 13,317 units that year, the majority from Chinese manufacturers Unitree and AgiBot. The Western players — Figure, Tesla, Boston Dynamics, Apptronik, and 1X — have collectively produced a handful of verified production deployments and a much larger number of claims. This article separates them.

The Benchmark: What “Production” Looks Like

Figure AI’s BMW Spartanburg trial is the most thoroughly documented Western commercial deployment to date. Two Figure 02 robots ran for 11 months on a single task: picking sheet-metal parts from racks and placing them on a welding fixture. Across 1,250 operating hours, they loaded more than 90,000 parts, covered approximately 1.2 million steps, and supported production of more than 30,000 BMW X3 vehicles. Cycle time held at 84 seconds per load (37 seconds for the placement itself). Accuracy stayed above 99%.

What the BMW trial also revealed: the forearm was the top hardware failure point, a predictable consequence of packaging three degrees of freedom and thermal constraints into a tight form factor. Figure has used that failure data to improve hardware reliability in subsequent designs. The trial ended in late 2025; BMW has now announced its first European humanoid deployment at its Leipzig plant, with a full pilot phase beginning summer 2026. The Leipzig application is high-voltage battery assembly — a more complex task than sheet-metal loading.

This is what production looks like in 2026: two robots, a narrow task envelope, extensive ramp time, and honest reporting of hardware failures. It is not a factory floor staffed with general-purpose humanoids. It is closer to a proof-of-concept with production-adjacent metrics.

The Tier-One Group: Verified Operational Deployments

Agility Robotics / Digit is the other confirmed production deployment. Digit has moved over 100,000 totes at GXO Logistics’ Flowery Branch facility. After 18 months of testing at Amazon’s Sumner facility, Agility reports a 98% task success rate — though this is a vendor-reported figure with no independent verification published, and “task success” for tote-moving describes a narrow, well-defined pick-and-place operation rather than general manipulation. The company has sold approximately 100 units as of 2025 and built a dedicated manufacturing facility in Salem, Oregon, with annual capacity exceeding 10,000 units.

There is a structural caveat: Digit operates in 30-minute intervals in practice at Amazon warehouses, despite a 90-minute maximum battery runtime. The IEEE Spectrum analysis identified battery life as “arguably the single most critical bottleneck” preventing industrial-scale deployment. No humanoid in commercial deployment today can run a full eight-hour shift without battery swap infrastructure. Figure 02 has a 2–3 hour active runtime. Boston Dynamics’ Atlas has four hours before it needs to swap — though it can do so autonomously in under three minutes.

Boston Dynamics / Atlas entered the commercial market formally at CES 2026, where the company unveiled its production-ready Atlas. All 2026 units are already committed: Hyundai’s Robotics Metaplant Application Center (RMAC) for automotive manufacturing and Google DeepMind for AI research. Specifications are the most capable of any commercially shipping humanoid: 56 degrees of freedom, 2.3-metre reach, 50 kg lift capacity, and four-hour battery life with autonomous swap. Hyundai, Atlas’s parent company, has announced $26 billion in US operations investment, including a robotics factory targeting 30,000 units per year capacity across all robot types. The deployment at RMAC represents the first genuinely scaled customer commitment in the Western market — but no commercial operating data, task performance metrics, or third-party deployment reports are yet available for Atlas.

The Tier-Two Group: Funded Pilots With Incomplete Data

Apptronik / Apollo has genuine commercial traction: confirmed partnerships with Mercedes-Benz, GXO Logistics, and Jabil, and a development lineage rooted in NASA Small Business Innovation Research (SBIR) contracts at Johnson Space Center. The company raised $520 million in February 2026 at a $5.5 billion valuation, with Google and Mercedes-Benz as lead investors, bringing total financing past $1 billion in under a year.

What Apptronik has not published: unit counts, task performance metrics, or facility-specific deployment data. The company describes Apollo as working on “real factory floors with some of the world’s biggest manufacturers,” which is consistent with active pilots. It is not consistent with verifiable production-scale deployment in the same tier as the BMW or GXO evidence. Apptronik should be watched as the most likely next entrant into the production tier — its funding and partner roster are stronger than most — but the data is not there yet.

1X Technologies (formerly Halodi) has taken an unusual path. The Norwegian-American company is targeting both consumer and industrial markets simultaneously. In December 2025, it announced a deal with EQT to deploy up to 10,000 NEO robots across EQT’s 300+ portfolio companies in manufacturing, warehousing, and logistics between 2026 and 2030. Consumer pre-orders opened at $20,000 for early access. The NVIDIA GTC 2026 presentation showed progress on simulation-based AI training infrastructure.

The EQT deal is notable for its scale — 10,000 units over four years across a diversified industrial portfolio. It is also notable for what it does not yet include: confirmed shipments, operating data, or even a confirmed first deployment. The deal is a framework, not a delivery record.

The Demo Tier: Where Claims Exceed Evidence

Tesla / Optimus is the most publicly followed and least independently verified entry in the field. Elon Musk has claimed thousands of Optimus units in productive factory use in 2025, tens of thousands in 2026, and a long-term target of 1 million units per year from a dedicated Gigafactory. On Tesla’s Q4 2025 earnings call on 28 January 2026, Musk acknowledged that no Optimus robots are doing “useful work” in factories — they are in the R&D phase, deployed primarily for data collection and algorithm iteration. Independent analysis suggests 2025 production ran in the hundreds of units, not thousands.

The Optimus V3 prototype is expected in 2026, with formal production beginning in summer 2026 at the converted Fremont facility. A dedicated Giga Texas factory targeting 10 million units per year has been announced. These are planning figures, not production figures. Tesla has not published task performance data, failure rates, or third-party verification of any factory deployment. Until it does, Optimus belongs in the demo tier — regardless of the production targets on the roadmap.

This is not unusual. The distance between a credible hardware roadmap and a production deployment with verified metrics is where most humanoid companies currently sit. Tesla is the highest-profile example, but others — including Sanctuary AI, Fourier Intelligence, and several pre-revenue entrants — fall into the same demo tier for the same reason: compelling prototypes, no published production data.

The China Variable

Any honest accounting of humanoid deployment in 2026 requires noting that Western companies are not the volume leaders. TechCrunch reported in February 2026 that Unitree alone shipped approximately 36 times more units in 2025 than US competitors Figure and Tesla combined. Chinese players Unitree and AgiBot collectively accounted for the majority of the ~13,317 global humanoid shipments in 2025.

The Chinese advantage is structural — the following factors are editorial analysis rather than individually sourced claims: a more integrated hardware supply chain, lower per-unit manufacturing cost, government policy support, and a domestic industrial base willing to absorb early-stage deployments. Western companies are competing on AI sophistication and premium customer relationships. Whether that trade-off holds as Chinese AI capabilities improve is an open question, but the volume gap is real and growing.

The Three Questions That Matter

For a chief technology officer (CTO) or plant engineering team evaluating humanoid procurement over the next 12–24 months, the evidence reduces to three questions.

What task? Every verified production deployment to date is narrow: sheet-metal loading, tote moving, parts kitting. No humanoid in commercial deployment is doing unstructured general labour. If the target task is not well-defined and repeatable, the robot is not ready.

What is the true duty cycle? Vendor specs describe peak capability. Production reality for current humanoids is 30–90 minutes of operation between charges or swaps. Eight-hour shift coverage requires infrastructure investment in charging stations or battery swap systems that vendors do not price into the robot cost.

What does the pilot agreement require? The BMW/Figure arrangement ran for 11 months with two robots on one task before BMW expanded to Leipzig. That ramp is realistic. Agreements that promise broad deployment within months are not — yet.

What to Watch

The BMW Leipzig deployment (full operation from summer 2026) will produce the first verified data on whether Figure can replicate its Spartanburg results at a different plant, with a more complex task, under European operating conditions. That is the cleanest near-term signal for the sector. It connects directly to the broader pattern of AI-enabled hardware deployment documented in our analysis of silicon quantum computing — where hardware milestones also lag claims by 12–18 months.

Boston Dynamics’ Atlas deployment at Hyundai RMAC will reveal whether a vertically integrated manufacturer-customer relationship accelerates the production learning curve. Hyundai’s $26 billion US investment makes them the most committed industrial patron in the Western humanoid market.

On Tesla, watch the Q2 2026 earnings call. If Optimus has moved from the “data collection” framing to “units doing measurable work” with any published metrics, that signals a real inflection. If the language is still forward-pointing production targets, it has not.

For Apptronik and 1X, the question is when they publish the first task-level performance data from a named deployment. Funding and partnerships are necessary conditions. They are not sufficient. The same standard applies to neuromorphic computing and other hardware categories where lab-stage benchmarks still run well ahead of verified commercial results.

The industry is moving. The gap between what is deployed and what is claimed is closing. But in 2026, it has not closed yet.

This article was produced with AI assistance and reviewed by the editorial team.