Humanoid robots stopped being a demo-stage curiosity this year, and the shipment numbers now prove it. New IDC data shows global humanoid shipments approached 25,000 units in the first half of 2026, a 432.1% jump from the same period a year earlier, with China’s Agibot taking the top spot.

The figures, published in early October, are the clearest snapshot yet of who is actually putting humanoids into the world, and where.

Agibot Takes the Lead

According to IDC, Agibot shipped more than 8,600 humanoid robots in the first six months of 2026. That works out to roughly 35% of the global market and more than 45% of the Chinese domestic market. IDC also ranks the company first in the industry by both shipment volume and revenue scale.

Agibot’s range spans four families: the full-size Yuanzheng humanoid, the half-size Lingxi, the wheeled Jingling, and the Kutuo quadruped. That breadth appears to be helping it sell into very different customers at once.

China Dominates the Count

China accounted for more than 19,000 of those units, about 77.9% of the global total, per IDC. Forbes, citing separate figures, put China’s share even higher, at around 95%, which shows how much the answer depends on how a “humanoid” is defined and counted.

A second tracker, Counterpoint Research, reached a somewhat lower total of more than 22,000 units, up nearly 300% year over year. In its ranking Agibot also led, at about 9,700 units and over 43% share, followed by Unitree with more than 7,000 units (31%), then Galbot, UBTech and Leju Robotics. The top five vendors, all Chinese, held 86% of shipments. The methodologies differ, but the direction is identical: a fast-growing market led by Chinese manufacturers.

Where the Robots Are Actually Working

Counterpoint’s breakdown is a useful reality check. Entertainment, performance, data production and research still account for more than 60% of shipments. Service and guidance roles make up about 19%, intelligent manufacturing has risen to 13%, and warehousing and logistics to 5%.

Agibot’s own deployments illustrate the spread:

• Tourism: more than 300 Agibot robots have worked at the Chimelong Space Shuttle park in Hengqin since September 24, handling performances and educational tours.
• Retail: 100 Lingxi X2 robots act as shopping guides in ASDA stores, working more than 10 hours a day.
• Industry: a Jingling G2 performs quality inspection on a tablet production line at Longcheer Technology’s Nanchang factory.

From Hardware Specs to Embodied AI

IDC says competition is shifting away from individual hardware specifications toward broader embodied AI capabilities. Counterpoint makes a similar point, expecting a race around closed-loop systems that combine models, vertical deployments, data pipelines and rapid iteration. Its forecast calls for more than 50,000 humanoid shipments in 2026, while IDC projects global volumes above 750,000 units by 2030. UBTech has already lifted its 2026 Walker S guidance to 5,000 units, and Agibot’s GO-3 model is slated to launch this autumn.

It is worth keeping the caveats in view. Many of these units are rented for shows, used for research or training data collection, or deployed as guides, not doing heavy industrial work. Western players such as Tesla and Figure are not featured in these rankings, and trackers disagree on totals.

What It Means for the Humanoid Race

The takeaway is that scale has arrived first in China, and it has arrived through volume, variety and low-risk use cases rather than a single breakthrough factory robot. For the rest of the industry, the next test is turning those thousands of deployed machines into steady, paid, productive work. If the second half of 2026 brings the manufacturing and logistics share up as Counterpoint expects, 2026 may be remembered as the year humanoids began shipping like a real product category. InteliDroid will keep tracking who is shipping, and who is working.

Boston Dynamics has spent nearly a decade proving Atlas can run, jump and backflip. Now it is betting the robot’s commercial future on something far less flashy: a hand with only four fingers.

On October 1, the Hyundai-owned company unveiled a redesigned Atlas hand with 13 degrees of freedom, nearly double the seven of its predecessor, and deliberately no pinky. In demo footage, Atlas loads a bit into a power drill, drives it into wood, tightens a small nut by fingertip and rotates two golf balls independently inside a single palm. The message is clear: the walking problem is solved, and the hand is where the humanoid race will be won or lost.

Why Boston Dynamics dropped the pinky

Almost every competing humanoid program has converged on five-fingered hands that mimic human anatomy as closely as possible. Boston Dynamics went the other way after what sounds like a remarkably low-tech experiment. Chief product and technology officer Zachary Jackowski asked the Atlas team to tape their pinky and ring fingers together for a day and report what they could no longer do. The answer, for most practical tasks, was very little.

Mechanical engineer Dylan Thrush explained the trade-off to The Robot Report: a fifth finger would add three more degrees of freedom, more mass, more power draw and more components to fail, without unlocking enough additional grasps to justify the cost. The four-finger layout still delivers the grips a factory floor actually demands, including:

• Pinch grasps between the thumb and any individual finger
• Tripod grasps across three fingers for holding irregular parts
• Sliding the thumb along and across every other finger for in-hand reorientation
• Triggered tool grasps for drills, torque drivers, grinders, nail guns and welding torches

The thumb alone carries four degrees of freedom and can oppose each of the three remaining fingers, while the hand itself is sized like a large human hand so Atlas can fit into workstations and use tools that were never designed for a robot.

Built for simulation first

The most interesting part of the design is not what you can see but what it was optimized for. Boston Dynamics says the hand was built from the ground up for sim-to-real reinforcement learning. Every joint uses a single type of directly actuated, fully encapsulated actuator with no cables crossing the joints, the same philosophy as the rest of the Atlas body. The result is a mechanism that can be simulated with high dynamic fidelity, which in turn lets the company train control policies against randomized friction, torque profiles, object geometries and disturbances before they ever touch hardware.

Alberto Rodriguez, director of robot behavior for Atlas, framed the design as a balance of competing goals rather than a search for a single right answer. Human demonstration data, captured through wearable interfaces, is good at teaching a robot what a task looks like. But fast, dexterous manipulation also depends on high-rate force control that wearables cannot capture. That gap is why Boston Dynamics believes reinforcement learning in simulation is essential, and why the hand’s backdrivable actuators and dense pressure sensors across the fingertips and palm matter so much. Early results, the company says, show promising transfer of dynamic behaviors trained purely in simulation onto the physical robot.

A hand with a job waiting for it

This is not a research curiosity in search of an application. The announcement landed one week after Boston Dynamics opened its Robotics Metaplant Application Center inside Hyundai Motor Group Metaplant America near Savannah, Georgia. Atlas units there are learning to sort and sequence automotive parts before they reach the assembly line, with component assembly targeted for 2030.

Hyundai and Kia have committed to deploying 25,000 Atlas robots across their global plants over the coming years, and Hyundai has said it intends to build a U.S. factory capable of producing up to 30,000 robots annually by 2028. Since Hyundai took full ownership of Boston Dynamics earlier this year, the Atlas pipeline is less a partnership than an internal supply-chain decision. The emphasis on manufacturability and repairability, a single actuator type, no fragile cables, a design meant to be serviced by a line technician, tells you exactly who this hand was built for.

The hand is the new bottleneck

Step back and a pattern emerges across the industry. Tesla’s Optimus Gen 3 is reportedly chasing 22 degrees of freedom per hand with tendon-driven actuators moved into the forearm. Figure 03 pairs palm cameras with fingertip sensors sensitive to a few grams of pressure. Unitree just put its 22-joint Dex5-S hand on sale for $6,500. Three very different companies, with very different philosophies, have all reached the same conclusion: the legs are good enough, and the hand is what stands between a demo and a shift.

Boston Dynamics’ answer is contrarian. Instead of maximizing degrees of freedom, it is minimizing complexity while keeping just enough dexterity to use real tools. Whether four fingers beat five will be decided not on a stage but on a Georgia factory floor over the next few years. For a field that has been measured by backflips for far too long, that is exactly the right place to settle the question.

While the humanoid headlines this month have been dominated by robots moving into people’s living rooms, a Japanese company just made a quieter—but arguably more strategic—move. On September 16, ugo Inc. unveiled ugo Nova, a semi-humanoid robot designed, engineered, and manufactured entirely in Japan and aimed squarely at one of robotics’ hardest bottlenecks: data. Rather than chasing viral backflips, Nova is built to do ordinary work while quietly recording every motion, turning human labor into fuel for the next generation of physical AI.

Meet Nova: A Semi-Humanoid Built for Work, Not Show

Nova is not a bipedal walker. It pairs a wheeled mobile base with a humanlike upper body: two seven-degree-of-freedom arms, a four-degree-of-freedom waist, and 22 degrees of freedom across the system. Standing 162 centimeters tall and weighing 90 kilograms, it is sized for the same doorways, workbenches, and counters that people use every day.

That “semi-humanoid” choice is deliberate. By skipping the enormous engineering cost and instability of dynamic legged locomotion, ugo can focus its budget on precise, dexterous manipulation and reliable, stable operation in real workplaces. The result is a platform that trades athletic spectacle for something more useful in the near term: a robot that can actually show up and get things done.

A Robot Designed to Be a Data Engine

Nova’s real purpose becomes clear once you look at how it is instrumented. Cameras are mounted as standard on both the head and the hands, capturing the visual data needed to train physical AI from the moment the robot is set up. Every task it performs can be converted directly into high-quality learning data for building robot foundation models.

Crucially, Nova is built for teleoperation. Operators can drive it using bilateral controllers and VR controllers, with arm control running at a high 100 Hz cycle so that movements are recorded smoothly and with high precision. A human guides the robot through a task—drilling, fastening, handling parts—and the system captures that demonstration as clean, structured training data. In its public demonstration, Nova performed a drilling task under remote control, exactly the kind of repetitive, physical job that is hard to simulate but easy to teach by example.

ugo announced Nova alongside a companion offering it calls the ugo Physical AI Fabric, a package that spans everything from data collection to foundation-model development. The message is that Nova is not just a robot, but one end of a full pipeline connecting real-world demonstrations to trained AI and, ultimately, to commercial deployment.

Why “Made in Japan” Matters Here

Nova arrives as part of a broader national push. The robot is being proposed by the AI Robot Association (AIRoA) under a project commissioned by NEDO, Japan’s New Energy and Industrial Technology Development Organization, to develop a domestically produced general-purpose robot. It was revealed at an event where prototypes from several Japanese companies were shown together—a coordinated signal that Japan intends to build homegrown hardware for the physical AI era rather than depend entirely on imported platforms.

For a country facing labor shortages and an aging workforce, a domestic, general-purpose robot that can be trained on local tasks is more than a tech demo; it is industrial strategy. Controlling the hardware, the data pipeline, and the models keeps critical capability—and the sensitive operational data that comes with it—inside the country.

The Takeaway

With mass production targeted for 2027, ugo Nova won’t flood workplaces overnight. But it reflects a shift that InteliDroid has tracked all year: the humanoid race is increasingly a race for data. The winners may not be the robots that move most impressively, but the ones that learn fastest from real human work. Nova is a bet that the shortest path to capable, general-purpose robots runs through thousands of hours of ordinary tasks—recorded, structured, and fed back into ever-smarter models. It is a distinctly Japanese answer to the physical AI question, and one worth watching as 2027 approaches.