What Humanoid Robots Actually Do in 2026
Search for humanoid robot deployment figures and you will find confident numbers. Tens of thousands of units in the field. Thousands working production lines. Most of those figures do not come from the companies they describe, and several do not survive contact with corporate filings [1]. That's only half the story; we need to answer a simple question: What do these machines actually do right now? This post is an attempt to answer that with evidence, including where the evidence runs out.
First, a way to read any humanoid claim
Almost every disagreement about humanoids dissolves once you sort claims into tiers. A useful taxonomy, adapted from analysts who track this carefully [1][2]:
Demo. A video. Possibly teleoperated, usually edited, always the best take. Tells you a capability ceiling under ideal conditions and nothing about reliability.
Announced. A partnership, an LOI, a scheduled fleet. A scheduled fleet is announced until shipment and operation are confirmed [2]. Corporate ownership of a robotics company by an automaker is not the same as that automaker being a validated customer.
Piloted. Robots physically present in a customer facility, doing constrained work under supervision, usually with engineers nearby.
Verified operational. Measured hours, measured output, a commercial contract, and numbers the operating company itself published.
The last tier is small. It's also the only one worth planning against.
The deployments with real numbers behind them
Warehouse tote handling — Agility Robotics' Digit. The clearest commercial record in the category. Digit operates under a multiyear robots-as-a-service contract at a GXO fulfillment facility, where Agility reported moving more than 100,000 totes, with over a year of continuous full-time deployment [3][5]. The work is specific: unloading autonomous mobile robots and feeding totes onto conveyors at pack-out stations [5]. Agility has since added commercial relationships with Toyota Motor Manufacturing Canada and Mercado Libre [3], and Digit has also been tested in Amazon fulfillment centers for tote recycling, moving empty containers between areas [12].
Automotive parts handling — Figure at BMW. Figure AI completed an eleven-month pilot at BMW Group Plant Spartanburg using its Figure 02 robot [1]. The reported operating detail is the most granular published for any humanoid program: ten-hour shifts Monday through Friday, more than 1,250 runtime hours, and more than 90,000 sheet-metal parts loaded, contributing to the production of over 30,000 vehicles [4].
Automotive intralogistics — Apptronik's Apollo at Mercedes-Benz. Apollo is running material-handling work at Mercedes facilities, focused on transporting components to the production line and performing initial quality checks [5]. Mercedes has taken an equity stake, and contract manufacturer Jabil both builds the robots and tests them on sorting, kitting, inspection, and subassembly [3]. Analysts who tier claims carefully classify this as evaluation rather than scaled autonomous production [2], an important distinction that most coverage collapses.
Consumer electronics assembly — AgiBot in China. AgiBot's G2 is deployed on Longcheer Technology's consumer electronics production line doing production tasks, and the company reported its 15,000th robot built as of June 2026 [7].
Tesla Optimus. Widely assumed to be the leader and the hardest to verify. As of mid-July 2026, production at Fremont had not begun; Tesla's Q2 delivery report contained no Optimus figures, and Tesla has never published a production count [1]. Analysts tracking the company's public communications counted eight separate Optimus production milestones delayed or revised before their original date [1]. Where Optimus is deployed internally at Tesla facilities, it is described as primarily generating training data rather than productive output [5].
The pattern: narrow, structured, and supervised
Line up every verified deployment and the tasks rhyme. Moving totes. Loading parts. Transporting kits. Sorting. Machine tending. The proven categories across documented 2026 deployments are material handling, parts transfer, bin picking, kitting, and simple assembly support [5]. What isn't on that list matters as much. High-speed, high-precision work remains firmly with traditional fixed industrial arms [5]. On picking tasks, humanoids have been measured at roughly 70 to 85 percent of human speed in warehouse pilots [12]. The reason for the pattern is straightforward once you know where the difficulty lives. These are tasks with tolerant success criteria, structured environments, and low contact complexity. They're the tasks where a policy trained on limited data can be reliable enough to run unattended. Precision assembly is contact-rich, and contact-rich work is where learned policies degrade hardest — it's the sharpest edge of the sim-to-real gap [13]. Worth noting: the first adoption wave was never about replacing a whole worker [4]. It's about absorbing one repetitive material-movement task inside a workflow that still has people in it.
The shipment numbers, and why they contradict each other
If you want a clean illustration of the evidence problem, look at 2026 shipment estimates. Counterpoint reported more than 22,000 units globally in the first half of 2026, close to 300 percent year-over-year growth, and projects over 50,000 for the full year [7]. Smart Analytics Global put the first half at roughly 19,100 units and forecasts about 60,000 for the year [8]. Gasgoo's research institute suggests full-year shipments could approach 100,000 [8]. Chinese sources claim over 30,000 units shipped in China alone [7]. They also disagree about what the robots are for. Counterpoint indicates most units go to entertainment and research; Smart Analytics Global says industrial and commercial applications account for over 70 percent [7]. Two research firms, same market, opposite characterizations. Part of the explanation is a category of buyer that complicates every count: state-backed training centers in China purchasing robots specifically to collect data [7]. Those are real units shipped and real revenue booked, and they represent something other than end-market demand for productive work [8]. Two things are simultaneously true: the market is concentrated; AgiBot and Unitree together account for roughly 75 percent of global shipments [7], and Western pure plays hold the strongest verified deployment records [1]. Volume and validation are not the same axis.
Where humanoids are not working
The domestic humanoid is the most-marketed and least-delivered application. As one analysis of home humanoid claims put it, no public demonstration has shown a robot autonomously completing a full laundry cycle—sorting, treating stains, loading, drying, folding, and putting away—on novel laundry, uncut [10]. Early home systems are expected to operate in tidy, well-defined homes under tight safety rules with teleoperation backup, handling narrow chore sets [12]. The gap is reliability, not capability. Bessemer Venture Partners characterized the current moment as robotics' "GPT-2.5 moment," with real capability and real scaling laws, and a wide gap remaining to the 99.9 percent reliability that production deployment demands [10]. That last stretch is where most of the hype lives, and Rodney Brooks, who co-founded iRobot, has called the broader humanoid narrative fantasy thinking [10]. Capital is not the constraint. Humanoid startups raised $8.6 billion in 2026 alone, roughly 1.8 times all of 2025's total with the year half over [9]. Manufacturing capacity is not the constraint either — building tens of thousands of units is achievable in the near term [11].
The actual bottleneck is data and deployment
Here's the detail that gets buried, and it's the most important one in this post: several of the most sophisticated humanoid programs are currently being used to generate training data rather than to produce output. Tesla's internal Optimus fleet is described that way [5]. Apptronik's current platform is offered in configurations explicitly for real-world task-data collection [2]. An entire category of Chinese buyers exists to collect data [7]. The companies closest to this problem are spending their most expensive hardware on the data bottleneck, because that is what stands between a robot that can do one demonstrated task and a robot that can do the next task. We've written about why that bottleneck is so expensive: robotics has no internet-scale corpus to train on, and a five-thousand-demonstration dataset can cost $250,000 to $750,000 to produce by hand [14]. Every humanoid deployment above required its own version of that effort for its own narrow task. Which points at the real question for anyone evaluating this technology. The form factor is not the interesting variable. A bipedal robot and a fixed arm face the identical problem: how do they learn this specific task in this specific cell, and how do they keep working when the task changes? The answer is automated training data collection and a deployment and operations layer that improves policies from field data, regardless of what the machine looks like. That's why 3GEN builds the deployment platform rather than a humanoid. Bellwether is designed to turn robotic deployment from an engineering project into a repeatable software process. When general-purpose hardware does mature, the constraint moves entirely to who can deploy and improve it, and that constraint exists today, on the robots already on your floor.
What to do with this if you're evaluating humanoids
Ask for verified operating hours, not units shipped. Units shipped includes research buyers, entertainment, and data-collection purchases. Hours-per-unit in a customer facility is the number that means something [9].
Check whether the task is on the proven list. Material handling, transfer, kitting, machine tending, and simple assembly support are reasonable. Precision assembly, welding, and high-speed work are not, and a vendor suggesting otherwise is describing a roadmap [5].
Ask who published the number. If a deployment figure doesn't trace to the company it describes or its customer, treat it as unverified [1].
Ask what happens when the task changes. Every deployment above is narrow. The economics depend entirely on whether the second task costs what the first one did.
Frequently asked questions
What do humanoid robots actually do today? Almost entirely material handling in structured environments: moving totes in warehouses, loading parts on automotive lines, transporting kits, sorting and kitting. Work is narrow, supervised, and defined per deployment [5].
Which humanoid robots are actually deployed commercially? Agility's Digit has the clearest commercial record at GXO under a robots-as-a-service contract [3]. Figure completed a measured eleven-month pilot at BMW Spartanburg [1][4]. Apptronik's Apollo is in evaluation at Mercedes-Benz [2][5]. AgiBot's G2 is on a production line in China [7].
Are humanoid robots replacing workers? Not in any documented 2026 deployment. The pattern is absorbing a single repetitive material-movement task within a workflow that still involves people [4].
Will humanoids work in homes soon? Not in the general sense being marketed. Early home systems are expected to run narrow chore sets in controlled environments with teleoperation backup [12], and no public demonstration has shown a full household task completed autonomously start to finish on novel inputs [10].
Are humanoids better than traditional industrial robots? Different, not better. Their advantage is operating in environments built for people — stairs, existing equipment, human-scale workstations. Fixed arms remain faster, more precise, and more reliable at repetitive high-tolerance work [5].
When will humanoids be widely adopted in factories? Broad factory-scale adoption looks more like a 2030s development than a 2026 or 2027 one, even though the category is genuinely operational today [4].
The honest summary
Humanoid robots in 2026 are neither the imminent workforce replacement being marketed nor the perpetual vaporware the skeptics describe. They are a real technology crossing a real threshold in a small number of bounded use cases, with a verified record measured in hundreds of units rather than the tens of thousands frequently claimed. The interesting question isn't whether the hardware will improve; it will. It's who can teach it new tasks fast enough to matter, which is a data and deployment question that applies with equal force to every robot in your facility today. If you have a task on your floor you want automated, request a demo or reach us at human@3genrobotics.com.
Notes
Bellwether is in alpha. Public demo material illustrates the pipeline conceptually; screens, flow, and terminology are subject to change in the shipped product. Performance figures cited above are drawn from published third-party research and vendor documentation, not from 3GEN internal benchmarks. This is a fast-moving area, and figures were current as of August 2026. Deployment details for individual programs are drawn from secondary reporting and company statements rather than independently audited disclosures; where sources conflict—particularly on which robot generation was used at which facility and on 2026 shipment totals—we have cited the more conservative figure and noted the disagreement. Shipment estimates from Counterpoint, Smart Analytics Global, and Gasgoo differ substantially in both magnitude and application mix [7][8], and no reconciliation between them was available at the time of writing. 3GEN Robotics does not manufacture humanoid robots and has no commercial relationship with any company named in this post.
References
[1] Technology.org. "Humanoid Robots in 2026: What Is Actually Deployed." July 18, 2026. https://www.technology.org/2026/07/18/humanoid-robots-in-2026-what-is-actually-deployed/
[2] TheMimic. "Every Humanoid Robot Company in 2026: The Complete Guide." July 2026. https://themimic.io/articles/humanoid-robot-companies-2026
[3] The AI Insider. "The State of Humanoid Robotics in 2026: Trends, Challenges and Opportunities." August 21, 2026. https://theaiinsider.tech/2026/08/21/the-state-of-humanoid-robotics-in-2026-trends-challenges-and-opportunities/
[4] New Market Pitch. "Humanoid Robots in Factories: For When?" June 2026. https://newmarketpitch.com/blogs/news/humanoid-robotics-factories-when
[5] iFactory. "Humanoid & Quadruped Robots for Manufacturing Plants 2026: Complete Industry Deployment Guide." May 2026. https://ifactoryapp.com/industries/manufacturing-plant/humanoid-quadruped-robots-manufacturing-plant-2026-guide
[6] Technerdo. "Humanoid Robots in 2026: Market Leaders, Deployments, and What Comes Next." April 2026. https://www.technerdo.com/blog/humanoid-robots-market-2026
[7] Koetsier, John. "Humanoid Robot Shipments Up 300%: Up To 30,000 So Far In 2026." Forbes, August 20, 2026, citing Counterpoint Research and Smart Analytics Global. https://www.forbes.com/sites/johnkoetsier/2026/08/20/humanoid-robot-shipments-up-300-up-to-30000-so-far-in-2026/
[8] Gasgoo. "The 'Schism' in Humanoid Robotics: Hype Goes Viral, But Practical Application Has 'Short Legs.'" August 2026. https://autonews.gasgoo.com/articles/news/the-schism-in-humanoid-robotics-hype-goes-viral-but-practical-application-has-short-legs-2088516075968020481
[9] Value Add Pulse. "Humanoid Robots: Follow the Manufacturing Data, Not the Demos." August 2026. https://valueaddvc.com/pulse/humanoid-robot-manufacturing-data-analysis-2026
[10] Chourasia, Mehul. "The Truth About Humanoid Robots in 2026." May 2026, citing Bessemer Venture Partners and The Robot Report. https://medium.com/@mehul.chourasia28/the-truth-about-humanoid-robots-in-2026-3ae82e9061b1
[11] Ackerman, Evan. "Reality Is Ruining the Humanoid Robot Hype." IEEE Spectrum. https://spectrum.ieee.org/humanoid-robot-scaling
[12] Winss Solutions. "Innovative Humanoid Robots in 2025–2026: Reality or Hype?" December 2025. https://www.winssolutions.org/humanoid-robots-2025-2026-reality-hype/
[13] Aljalbout, Elie, et al. "The Reality Gap in Robotics: Challenges, Solutions, and Best Practices." Annual Review of Control, Robotics, and Autonomous Systems, Vol. 9, 2026. arXiv:2510.20808. https://arxiv.org/abs/2510.20808
[14] DataX Power. "Humanoid Robot Data Collection Costs: 2026 Real Benchmarks by Program Type." July 2026. https://www.dataxpower.com/blog/humanoid-robot-data-collection-cost