In 2026 the American humanoid industry stopped being a demo reel. Six companies now cover the entire stack, from actuators to whole machines to foundation models, and the capital density behind them is extraordinary. The six are Tesla, Figure, Boston Dynamics, Agility Robotics, Apptronik, and 1X, a Norwegian company that has since planted itself in the United States.
Reviews of this field usually turn into leaderboards. Who has the highest specification, the largest valuation, the biggest round. That misses the point. What actually separates these six is the answer each one gave to three questions. Who buys the robot, a factory or a household? Which technical route, electric or hydraulic, tendon drive or modular? And where does commercialisation land, inside your own plant first or with outside customers?
2026 delivered the turn. Several of these machines entered real production lines, and some of them started generating real revenue. That moves the sector from technical demonstration to commercial validation. Sorted by route, the six fall into three camps. The factory camp treats the robot as a productivity tool and includes Tesla, Figure, Boston Dynamics, Agility and Apptronik. The home camp has exactly one member, 1X. The platform camp sells the robot plus the foundation model as one capability, and both Figure and 1X are pushing hard there.
Tesla Optimus Gen 3: car manufacturing, transplanted
Tesla is doing the one thing nobody else can copy. It is moving the scale and vertical integration of car making into robots.
Optimus Gen 3 started mass production in January 2026 at the Fremont plant in California. That line used to build the Model S and Model X. Tesla killed both cars and converted the building into a robot manufacturing facility. On the first quarter earnings call Musk put it bluntly: Optimus contains roughly 10,000 separate parts, and the speed of the whole line is set by the unluckiest, slowest, dumbest component in that list.
The specification jump is real. Gen 3 weighs 57 kg against 73 kg for Gen 2, a 22 per cent reduction, at 173 cm tall. The hands carry the biggest upgrade. Each hand now has 22 degrees of freedom, double the 11 on Gen 2, with 25 actuators packed into each forearm and 50 across both hands. The drive imitates tendons, with motors in the forearm pulling the fingers, matching human anatomy. That design lets Gen 3 perform more than 3,000 fine manipulations against roughly 500 for Gen 2. Fingertips carry force feedback sensors, so it can hold an egg or a glass bottle without crushing it.
Compute comes from Tesla’s own AI5 chip, with about five times the memory bandwidth of the previous AI4. Sensing is eight autonomous-driving-grade cameras, pure vision, no lidar. It runs xAI’s Grok voice model locally. Dynamic payload is 20 kg, a squat lift reaches roughly 68 kg, and runtime exceeds eight hours.
Commercially, Tesla eats its own cooking. By March 2026 more than 1,000 Optimus units were already inside Fremont and the Texas gigafactory, sorting battery cells, moving parts and running basic quality checks. External sales wait until after 2027. The long-term price target is 20,000 to 30,000 dollars per unit, conditional on annual output reaching one million.
Tesla’s real advantage is not one chip or one hand. It is a flywheel. Spatial understanding trained on millions of miles of FSD data, amplified by the Cortex 2.0 supercomputer, then fed back by robots doing real work on real factory floors. A pure robotics startup cannot reproduce that loop.
Figure 03: the first verified return on investment
If you are allowed only one case that a factory has already validated, Figure 03 is the strongest available.
Figure was founded in 2022 by Brett Adcock. The latest round put valuation at 39 billion dollars on cumulative funding above 2.4 billion dollars, with OpenAI, Microsoft, NVIDIA and Bezos behind it. That investor list is itself a signal.
Figure 03 stands 1.70 m, weighs roughly 57 to 61 kg, carries 48 to 50 degrees of freedom, lifts 25 kg and runs five to six hours. The hands get the most praise. Each fingertip senses pressure changes at the 3-gram level, about the weight of a feather landing, and each palm holds a wide-angle low-latency camera that removes the close-range blind spot during a grasp.
The brain is Helix, Figure’s own vision, language and action model. Helix 02, released in January 2026, extends control from the upper body to the whole body across three coordinated layers. System 2 at the top is a 7-billion-parameter vision language model running at 7 to 9 Hz, deciding what to do next. System 1 in the middle is an 80-million-parameter visuomotor policy network running at 200 Hz, converting semantic goals into whole-body joint trajectories. System 0 at the bottom is a 10-million-parameter base control module running at 1 kHz, holding balance and contact.
Validation came from BMW. At the BMW plant in Spartanburg, South Carolina, Figure’s machines worked for 11 months across a line that produced more than 30,000 vehicles, at 99 per cent task accuracy. That is the industry’s first case of return on investment at scale inside a real car plant. In March 2026 Figure 03 also appeared at a White House AI education event, which carries obvious policy weight. On manufacturing, Figure’s own BotQ plant is planned for 12,000 units a year.
Boston Dynamics Atlas: the ceiling on movement
The new Atlas is the most athletic of the six. Boston Dynamics is not chasing the household first, and it is not rushing to prove return on investment. It is pushing raw capability to the limit.
On 5 January 2026 at CES, Boston Dynamics launched the production version of the all-electric Atlas and started building it at its Boston headquarters the same day. It stands 1.9 m, weighs about 89 kg, has 56 degrees of freedom with omnidirectional joints, a 2.3 m arm span, and lifts 50 kg.
The standout feature is self battery swap. The robot navigates to a charging station, pulls the old pack, fits a new one, and finishes in under three minutes, which makes uninterrupted operation possible. It works from minus 20 to 40 degrees Celsius and is waterproof.
This route walked from hydraulics to electric. In April 2024 Boston Dynamics retired the hydraulic Atlas and replaced it with a full electric redesign. After electrification, the hips, waist and neck all rotate 360 degrees, which humans cannot do. Actuators come from Hyundai Mobis, plugging into an automotive-grade supply chain.
All 2026 output is already locked to Hyundai’s RMAC plant and to Google DeepMind. Hyundai Group is the controlling shareholder and in June 2026 bought SoftBank’s remaining stake of roughly 9.65 per cent, then announced a 26 billion dollar US investment programme that includes a robot plant with capacity of 30,000 units a year. The DeepMind partnership plugs frontier foundation models into the machine to lift cognition.
Boston Dynamics does not survive on stunts. It survives on industrial reliability and continuous operation.
Agility Digit: the only one earning warehouse money
Of the six, Agility’s Digit is the only machine producing real revenue inside a real warehouse.
Digit was spun out of Oregon State University in 2015 and designed for logistics from birth. It stands about 1.55 to 1.75 m, weighs 42 to 65 kg, carries 16 to 18 kg, and runs four to eight hours. It does not chase generality. It does one thing, which is moving totes in a warehouse.
The commercial data is unusually concrete. By April 2026, Digit deployment at Amazon’s Spanaway fulfilment centre in Washington state reached 75 units, the largest single humanoid deployment to date, with more than 100,000 totes moved. Agility charges through a robotics-as-a-service model at roughly 25 dollars an hour, against an all-in US warehouse worker cost above 30 dollars an hour. The company wants operating cost down to 2 to 3 dollars an hour once scale arrives.
Customers extend beyond Amazon. Toyota Canada signed a service agreement in February 2026, and GXO and Latin American marketplace Mercado Libre are also users. On manufacturing, Agility runs RoboFab in Salem, Oregon, nameplate peak capacity 10,000 units a year, with 75 per cent of roughly 6,000 parts sourced inside the United States. Chief executive Peggy Johnson has committed to delivering the first collaborative-safe Digit by the end of 2026, a version that needs no safety fencing and drops itself to the floor when a person comes close.
Digit does not do backflips. It works steadily in the most boring environment available. Boring is where the cash flow lives.
Apptronik Apollo: the NASA-adjacent pragmatist
Apptronik is the quietest of the six and the most like an enterprise product. It came out of the Human Centered Robotics Lab at the University of Texas at Austin and carries NASA collaboration in its DNA.
Apollo stands 1.73 m, weighs 73 kg, lifts 25 kg, runs four hours, and has 32 to 36 degrees of freedom. Its core play is modularity. Fit legs and it walks, fit wheels and it rolls, or bolt it to a fixed base for repetitive work. Batteries are hot swappable at four hours per pack, a swap takes under five minutes, and the machine can cover 22 hours a day. Hands swap too, so the robot picking boxes on Tuesday can load machine tools on Wednesday. The reason for modularity is entirely practical. It is aimed at commercial deployment.
The endorsements are heavy. Mercedes-Benz, GXO and Jabil are all pilot customers, and a strategic partnership with Google DeepMind supplies cognition through Gemini Robotics. In February 2026 Apptronik closed a 520 million dollar A-X round, taking total funding above 935 million dollars at a 5.5 billion dollar valuation, with John Deere, AT&T Ventures and the Qatar Investment Authority joining. In the Mercedes pilot, logistics efficiency improved by roughly 40 per cent.
The neatest move is robots building robots at Jabil. Newly assembled Apollo units run quality checks, sorting and delivery inside Jabil’s own workshop, get validated, then ship to customers, creating a flywheel of falling cost and rising scale.
1X NEO: the only one aiming at your living room
The first five are all looking at factories. Only 1X is looking at the home.
1X started in Norway in 2014 as Halodi Robotics, pivoted to consumer humanoids in 2022, and now runs dual bases in the United States and Norway, with the OpenAI Startup Fund as an early lead investor.
NEO Gamma weighs 30 kg, stands 165 to 168 cm, has 22 degrees of freedom in the hands, lifts around 68 kg, and runs at 22 decibels, quieter than a refrigerator.
The technical route differs from everyone else. 1X uses its own tendon drive system that imitates human tendons and muscle, wrapped in a custom 3D lattice polymer soft shell for safety and cushioning. The brain is Redwood AI, a 160-million-parameter vision, language and action model running on an onboard NVIDIA Jetson Thor at 2,070 TFLOPS, with no cloud dependency. The company also trained a world model for simulation.
Pricing is 20,000 dollars outright or 499 dollars a month, with US deliveries starting in 2026. Preorders opened in October 2025 and the first year allocation of 10,000 units sold out in five days. The NEO plant in Hayward, California is ramping towards 10,000 units a year, and a San Carlos facility targets 100,000 a year by the end of 2027.
A Morgan Stanley analyst bought one and took it home. The verdict was cold. At this stage it looks more like a financing strategy than a genuinely useful household tool, and early units still need remote human assistance.
That is a harsh review, and it points at a real problem. The home is harder than the factory because the environment is open, tasks are vague, and the safety bar is high. Yet the question stands. Is the household endgame really easier to win than the factory?
What the split actually tells you
Six American challengers, three customer types, three technical routes. In the factory camp, Tesla runs on scale and its data flywheel, Figure on validation and AI, Boston Dynamics on athletic capability and reliability, Agility on actual revenue, Apptronik on modularity and enterprise relationships. Only 1X faces the household directly, betting on soft materials and subscription pricing.
Factories are the table. The home is the endgame.
Editor’s note: This is an adapted translation of the original OFweek report. It has been trimmed and restructured for readability for an international business audience.