Six players, three routes
In 2026 the American humanoid-robot industry began to protect itself. The United States now has at least six serious robot companies: Tesla, Figure, Boston Dynamics, Agility, Apptronik, plus 1X, which was born in Norway and landed in the United States. Six companies, six completely different technical routes. A roundup can look like a leaderboard of parameters, valuations and funding, but what matters is the different choices each makes on the questions that decide the business: who the robot is sold to, factory or home; which technical route, electric or hydraulic, tendon-driven or modular; and where commercialisation lands, in its own factory first or sold to outside customers.
A turning point came in 2026. Several companies’ machines entered real production lines and some booked real revenue, pushing the industry from the demo stage to commercial validation. America’s players cover the full chain from actuators to whole robots to AI foundation models, and capital density is extremely high. The routes group into three camps: the factory camp treats the robot as a productivity tool; the home camp, today only 1X, makes a soft companion humans will accept; and the platform camp sells robot plus foundation model as a whole, with Figure and 1X both pushing there.
Tesla Optimus Gen 3: moving car-making into the robot
Tesla moved the scale and vertical integration of car-making into the robot, something others cannot copy. Optimus Gen 3 began mass production in January 2026 at the Fremont factory in California, on a line that used to build the Model S and Model X; Tesla stopped those two cars and converted the building into robot manufacturing. Musk was blunt on the first-quarter call: Optimus has about 10,000 separate parts, and line speed depends on the unluckiest, slowest, dumbest part.
Gen 3 is 57 kg, 22 per cent lighter than the 73 kg Gen 2, and 173 cm tall. The hand is the biggest upgrade: 22 degrees of freedom per hand, double the Gen 2’s 11, with 50 actuators in the forearms using tendon-style drive that mirrors human hand structure. It can perform more than 3,000 fine manipulations against about 500 for Gen 2, with fingertip force-feedback sensors that can hold an egg or a glass bottle without crushing it. Compute comes from Tesla’s self-designed AI5 chip with about five times the memory bandwidth of AI4. It uses eight autonomous-driving-grade cameras, pure vision, no lidar, and runs xAI’s Grok voice model on board. Dynamic load is 20 kg, it can squat-lift about 68 kg, and battery lasts over 8 hours.
Commercially Tesla uses it itself first. By March 2026 more than 1,000 Optimus units were sorting batteries, moving parts and doing basic inspection at Fremont and the Texas gigafactory. External sales wait until after 2027, with a long-term target price of US$20,000 to US$30,000 once annual output reaches one million units. Tesla’s real edge is a flywheel: millions of FSD training miles teach spatial understanding, Cortex 2.0 supercomputers scale the training, and factory robots feed real manipulation data back. That loop a pure robot startup cannot copy.
Figure 03: the first validated return on investment
If only one factory-proven case can be picked, Figure 03 is the most convincing. Founded in 2022 by Brett Adcock, its latest round valued it at US$39 billion with more than US$2.4 billion raised, backed by OpenAI, Microsoft, Nvidia and Bezos. That capital line-up is itself a signal. Figure 03 is 1.70 m, about 57 to 61 kg, 48 to 50 degrees of freedom, 25 kg load, 5 to 6 hours of battery. Its celebrated hand senses pressure changes at the 3-gram level, a feather’s weight, and each palm carries a wide-angle low-latency camera to kill the close-range blind spot. Its brain is Helix, Figure’s own vision-language-action model.
Helix 02, released January 2026, extended control from the upper body to the whole body in three coordinated layers: System 2 is a 7-billion-parameter vision-language model running at 7 to 9 Hz to decide what to do next; System 1 is an 80-million-parameter visuomotor policy at 200 Hz turning semantic goals into whole-body joint trajectories; System 0 is a 10-million-parameter base control module at 1 kHz holding balance and contact. At BMW’s Spartanburg plant in the United States, Figure’s machines worked 11 months across more than 30,000 vehicles at 99 per cent task accuracy, the first scaled ROI case in a real auto factory. In March 2026 Figure 03 also appeared at a White House AI education event, a heavy policy signal. Its BotQ factory is planned at 12,000 units a year.
Boston Dynamics Atlas: the ceiling of motion
Boston Dynamics’ new Atlas is the most capable mover of the six. It does not rush the home market or prove ROI first; it pushes what a robot can do and carry to the limit. On 5 January 2026 at CES it launched the mass-production version of the all-electric Atlas, starting production that day at its Boston headquarters. It is 1.9 m, about 89 kg, 56 degrees of freedom, joints rotating in all directions, 2.3 m arm span, 50 kg lift. Most distinctive is self-battery-swap: the robot navigates to a charger, pulls the old battery, slots in the new one, all in under 3 minutes, for uninterrupted work. It runs from minus 20 to 40 degrees Celsius and is waterproof. The route moved from hydraulic, retiring the hydraulic Atlas in April 2024, to a full electric redesign; after electrification the hips, waist and neck rotate 360 degrees. Actuators come from Hyundai Mobis on an auto-grade supply chain. Its 2026 capacity is locked entirely to Hyundai’s RMAC factory and Google DeepMind. Hyundai is the controlling shareholder and in June 2026 bought SoftBank’s remaining 9.65 per cent and announced a US$26 billion US investment plan including a 30,000-unit-a-year robot factory. Boston Dynamics does not show off; it relies on industrial-grade reliability and uninterrupted operation.
Agility Digit: the only one truly earning in warehouses
Of the six, Agility’s Digit is the only one generating real revenue in real warehouses. Spun out of Oregon State University in 2015 and built for logistics, it is 1.55 to 1.75 m, 42 to 65 kg, 16 to 18 kg load, 4 to 8 hours, and does one thing: move totes in warehouses. By April 2026 Digit reached 75 units at Amazon’s Spanaway, Washington fulfilment centre, the largest single humanoid deployment to date, having moved more than 100,000 totes. Agility charges a robots-as-a-service rate of about US$25 an hour against US$30-plus for a US warehouse worker, with a goal of US$2 to US$3 an hour at scale. In February 2026 Toyota Canada signed a RaaS deal; GXO and Latin American e-commerce firm Mercado Libre use it too. Its RoboFab plant in Salem, Oregon peaks at 10,000 units a year with about 75 per cent US-made parts. CEO Peggy Johnson promised the first collaborative-safe Digit, which drops to the floor to avoid a person instead of needing a cage, by the end of 2026. Digit does not backflip; it works steadily in the most boring places, and boring places are where the cash flow really is.
Apptronik Apollo: the pragmatic NASA-blooded player
Apptronik is the most low-key and the most enterprise-product-like. It grew from the Human Centered Robotics lab at the University of Texas at Austin with NASA collaboration in its DNA. Apollo is 1.73 m, 73 kg, 25 kg load, 4 hours, 32 to 36 degrees of freedom. Its core play is modularity: legs to walk, wheels to run, or a fixed base for repeat operations; hot-swap batteries of 4 hours each, swapped in under 5 minutes, sustaining 22 hours a day; swappable hands, so Tuesday’s machine bins parts and Wednesday’s tends a machine tool. Backers include Mercedes, GXO and Jabil, and it works with Google DeepMind’s Gemini Robotics for cognition. In February 2026 it closed a US$520 million A-X round, total funding over US$935 million, valuation US$5.5 billion, with new investors John Deere, AT&T Ventures and the Qatar Investment Authority. In Mercedes pilots logistics efficiency rose about 40 per cent. Most elegant is the Jabil factory where robots build robots, a cost-down scale-up flywheel. Apptronik does not chase traffic; it chases factory contracts.
1X NEO: the only soft home companion
The first five watch the factory; only 1X targets the home. Founded in Norway in 2014 as Halodi Robotics, it turned to consumer humanoids in 2022 and is now dual-based in the US and Norway, with OpenAI’s startup fund as an early lead investor. Its NEO Gamma is 30 kg, 165 to 168 cm, 22 degrees of freedom in the hand, lifts about 68 kg, and runs at just 22 decibels, quieter than a fridge. It uses a self-developed tendon system mimicking human tendons and muscles under a custom 3D lattice polymer soft shell for safety and cushioning. Its brain is Redwood AI, a 160-million-parameter vision-language-action model on an on-board Nvidia Jetson Thor at 2,070 TFLOPS, no cloud, with a world model for simulation. Price is US$20,000 to buy or US$499 a month to subscribe, with US deliveries starting in 2026; preorders opened in October 2025 and the first 10,000 sold out in five days. The Hayward, California NEO factory is ramping to 10,000 a year and the San Carlos plant targets 100,000 a year by late 2027. A Morgan Stanley analyst bought one to test and concluded it is more a financing strategy than a truly useful house tool, still needing human remote assistance, which points at a real problem: the home is harder than the factory, with open environments, vague tasks and high safety bars.
The board, the endgame
Six American robot contenders, three kinds of customer, three technical routes. The factory camp: Tesla by scale and flywheel, Figure by validation and AI, Boston Dynamics by motion and reliability, Agility by real revenue, Apptronik by modularity and enterprise ties. Only 1X faces the home fully, betting the endgame on softness and subscription. The factory is the poker table; the home is the final board.
Editor’s note: This is an adapted translation of the original OFweek Robotics report. It has been trimmed and restructured for readability for an international business audience. The full original (in Chinese) is at https://mp.ofweek.com/robot/a556714295647.