AI in Robotics 2026: Companies, Models, What’s Shipping

A robot from a South Carolina startup built 30,000 cars at a BMW plant in 2025 with 99% accuracy. Not a fixed industrial arm bolted to the floor — a humanoid robot, walking the factory line, picking parts, learning the job. The robot is Figure 02. The startup is Figure AI, which has raised $1.9 billion and was last valued at $39 billion. The deployment cut BMW’s per-unit assembly cost by roughly 40% on the specific tasks Figure handled. As of mid-2026, Boston Dynamics’ Atlas is in production at Hyundai factories, Agility’s Digit is moving over 100,000 totes a year at GXO Logistics’ Spanx warehouse, and Apptronik’s Apollo is testing at Mercedes-Benz plants in Europe.

This is not a science-fair preview. This is the first year humanoid robots became a category you can buy. The robots are not yet at your local Costco, and most still trip on uneven ground, can’t reliably handle unfamiliar tools, and need humans to recover when something goes wrong. But the foundation-model breakthrough that powered ChatGPT has also rewritten what robots can learn — and the resulting wave of capital, factory deployments, and shipping units is the most consequential thing happening in the physical world this decade. This guide walks through what’s real, what’s still demo, who the players are, and where the technology is going.

What is AI in robotics — in plain language?

Industrial robots have existed since the 1960s. The Unimate arm that Volkswagen and General Motors used to weld car bodies was a robot. So is the welding cell in any modern auto plant, the painting line in an aerospace factory, and the pick-and-place arm in a phone-assembly line in Vietnam. The International Federation of Robotics counts roughly 4.3 million industrial robots installed worldwide as of 2024. None of those are what people mean when they say “AI robotics.”

The new wave is different along three specific axes. First, the robots have general bodies — usually humanoid form, two arms with hands, two legs (or wheels), eyes (cameras), a torso. This means one robot can theoretically do many tasks instead of one robot per fixed job. Second, the robots are trained, not programmed. Instead of an engineer writing a script for “pick up bolt, rotate 90 degrees, insert,” the robot learns from human demonstrations or from massive datasets of how objects move in the physical world. Third, the brain is a foundation model — software trained on huge quantities of data the way large language models like ChatGPT are trained on text, but here trained on vision, action, and language together.

Put plainly: the old robots were good employees who only ever did one job and needed a programmer to retrain them. The new robots are more like apprentices — you can show them a new task, and after enough demonstrations they pick it up. They are not yet as good as a human worker. But the rate of improvement is fast and the cost trajectory is dropping in ways the field has never seen before.

Why is 2026 the year humanoid robots became a real category?

Four things converged. The price of electric actuators — the motors that move robot joints — dropped enough that a full humanoid robot now costs in the tens of thousands of dollars to build instead of hundreds of thousands. Boston Dynamics, Hyundai Mobis, and Chinese suppliers all pushed actuator economics forward simultaneously. Second, foundation models trained on combined video, language, and action data crossed a capability threshold around 2023-2024 — Google’s RT-2, Physical Intelligence’s Pi Zero, Figure’s Helix, and Skild AI’s Brain are the named examples, but every major humanoid company now has a comparable in-house model.

Third, the data problem started yielding. A humanoid robot can learn by watching video of humans doing tasks, by being teleoperated through a job hundreds of times, or by training in physics simulators that produce billions of synthetic training examples. None of these were realistic five years ago. Fourth — and this is the boring but decisive one — large industrial customers like BMW, Mercedes-Benz, Toyota, Hyundai, Amazon, and GXO Logistics decided in 2024-2025 to pilot humanoid robots in production, generating revenue and feedback that no demo video ever could.

The combination produced the inflection point. Capital flooded in: Figure raised at $39B, Apptronik at $5B, Tesla committed multibillion-dollar production lines for Optimus, Chinese players like Unitree shipped a $16,000 humanoid (the G1) at consumer-electronics price points. Whether the projections of “millions of humanoid robots by 2030” hold up is uncertain — but the question is no longer whether the category exists.

Who are the major humanoid robotics companies in 2026?

Company	Robot	HQ	Latest funding	Real deployment
Figure AI	Figure 02 / 03	Sunnyvale, CA	$1.9B total, $39B valuation (Sept 2025)	BMW Spartanburg (30K+ cars built); BotQ factory online
Boston Dynamics	Atlas (electric)	Waltham, MA	Owned by Hyundai Motor Group	In production March 2026; shipping to Hyundai RMAC + Google DeepMind
Tesla	Optimus (V3 mid-2026)	Palo Alto, CA	Funded by Tesla Inc.	Internal Tesla factories; V3 production summer 2026, hundreds of units
Agility Robotics	Digit	Albany, OR	Backed by Amazon, DCVC, Playground Global	GXO (100K+ totes/yr), Amazon, Spanx, Toyota Mfg Canada
Apptronik	Apollo	Austin, TX	~$1B total raised, $5B valuation (Feb 2026)	Mercedes-Benz factory pilots, GXO Logistics R&D partnership
1X Technologies	NEO	Moss, Norway / Sunnyvale	Backed by OpenAI Startup Fund, EQT Ventures	Pre-orders open for home version; commercial pilots
Sanctuary AI	Phoenix	Vancouver, BC	Backed by Magna, Bell, BDC	Pilot programs; not yet at production scale
Unitree Robotics	G1 / H1	Hangzhou, China	Private; Chinese consumer hardware giant	G1 ships at ~$16K to research labs globally; H1 industrial pilots
Galbot / UBTech / Fourier	Various	China	State + private backing	Chinese factory pilots; not yet broadly exported

Three of these names belong in any first conversation about the field. Figure is the most-capitalized pure-play humanoid company in the world. Boston Dynamics is the technical credibility leader and the first to formally enter production for enterprise customers. Tesla is the wildcard with by far the largest potential production scale — though the company is famous for missed timelines, and as of mid-2026 the actual V3 hardware has not yet been revealed. The other names are real players with real deployments, but smaller scale.

Where are humanoid robots actually working in 2026?

This is the question that separates news from marketing. Here’s what’s verifiable as of mid-2026:

Deployment	Robot	What it does	Status
BMW Spartanburg (USA)	Figure 02	Sheet-metal placement, body-in-white tasks	30,000+ vehicles built; production-scale
GXO Logistics Flowery Branch (Spanx)	Agility Digit	Tote-moving, palletization assistance	100,000+ totes moved (fall 2025); multi-year RaaS contract signed
Mercedes-Benz factories (Europe)	Apptronik Apollo	Material handling, repetitive assembly tasks	Pilot phase; expanding
Toyota Motor Mfg Canada	Agility Digit	Logistics inside the plant	Post-pilot rollout agreement signed
Amazon (multiple sites)	Agility Digit + others	Tote handling, internal logistics	Multi-site pilot
Hyundai RMAC + Google DeepMind	Boston Dynamics Atlas	Factory automation; foundation-model research	First commercial Atlas shipments March 2026
BMW Plant Leipzig (Germany)	Hexagon AEON (wheeled humanoid)	European pilot — wheeled, not biped	Starting summer 2026; Figure 03 evaluation ongoing for other use cases
Tesla Fremont	Optimus (early generation)	Internal Tesla factory tasks; battery cell handling	Limited internal deployment; V3 production ramps summer 2026

What’s notably not on this list: hospitals, restaurants, retail floors, schools, or homes. Humanoid robots in 2026 are working in factories and warehouses — environments that are structured, repetitive, and forgiving of robot limitations. They are not yet working in unstructured human environments. That gap is the difference between the next two years of progress and the next ten years.

What are robotics foundation models, in plain English?

A foundation model is the brain of a modern robot. It is software trained on enormous amounts of data — videos of humans doing tasks, robot demonstrations, simulator-generated experiences, and text descriptions — that learns general patterns about how to perceive the world and how to act in it. Once trained, you can give the model a task in natural language (“pick up the red mug and put it in the sink”) and it figures out the sequence of movements needed.

The shorthand for this category is Vision-Language-Action models, or VLA. They take in what the robot sees (vision), what the human or system asks for (language), and produce what the robot should do (action). This is fundamentally different from how robots worked five years ago. The old approach was: an engineer wrote a program for every task. The new approach is: the model has seen enough that it can generalize to new tasks without being explicitly programmed.

Model	From	What’s notable
RT-2 / RT-X	Google DeepMind	First major demonstration that web-scale vision-language data transfers usefully to robot control
Pi Zero (π0)	Physical Intelligence	Open-style cross-embodiment model designed to work across different robot bodies
Helix	Figure AI	Figure’s in-house VLA; powers BMW deployment
Atlas brain	Boston Dynamics + Google DeepMind	Joint development using DeepMind foundation models
Skild Brain	Skild AI	“Robotic foundation model” designed to be hardware-agnostic; Vinod Khosla, SoftBank backed
Optimus brain	Tesla	Closely-held; assumed to share architecture with FSD/Dojo training infrastructure
Sanctuary Carbon	Sanctuary AI	“Cognitive architecture” approach combining reasoning + control

The race in 2026 is whether one foundation model can generalize across many robot bodies (Pi Zero’s bet, Skild’s bet) or whether each humanoid company needs its own purpose-built model that’s tightly integrated with its hardware (Figure’s bet, Tesla’s bet). The answer will shape who wins.

How are humanoid robots different from regular industrial robots?

The honest answer is that industrial robots are dramatically better at the tasks they’re designed for, and humanoid robots are dramatically better at the rest. A welding robot in a Ford plant has been welding the same seam at the same angle for fifteen years with no human intervention. It costs roughly $100,000 installed, runs essentially 24/7, and would be embarrassed to be compared to a humanoid robot for that specific task.

What humanoid robots do that industrial robots cannot: walk through a factory aisle without modifying the building, take instructions in natural language, learn a new task in hours instead of weeks of engineering integration, work on lines that were designed for human workers (no need to re-engineer the assembly station for robot arms), and physically reach into spaces designed for the human body.

Dimension	Traditional industrial robot	Humanoid robot
Cost	$25K–$200K depending on size	$30K–$150K (2026 estimates; rapidly dropping)
Setup time	Weeks to months of engineering	Hours to days of demonstration / teleop
Tasks per robot	1 (sometimes 2–3 with reconfiguration)	Many; reprogrammed in natural language
Works in human spaces	No — facility redesign required	Yes — that’s the point
Reliability	Very high; millions of cycles documented	Improving; not yet at industrial reliability
Best for	Repetitive single-purpose work	Flexible, multi-task work in mixed-human spaces

In most modern factories the answer is “both.” The welder stays. The new humanoid handles tote-moving, parts-staging, and the awkward in-between tasks that didn’t justify a dedicated robot but were always pulling humans away from higher-value work.

What’s still hard for humanoid robots in 2026?

The reality gap between demo videos and shipping robots is real. Things humanoid robots in 2026 are still demonstrably bad at:

• Unstructured environments. A robot that performs reliably in a clean factory aisle still fails in a real home with cats, throw rugs, and a stack of mail on the entry table.
• Dexterous manipulation. Picking up a sock. Putting a key in a lock. Tying a shoelace. The hands are improving fast (Tesla’s V3 22-DOF hand is the most-watched bet), but human-level dexterity is years away.
• Battery life. Atlas operates for ~4 hours; competitors are similar. Factories handle this with hot-swap battery routines. Homes won’t.
• Recovery from errors. When a humanoid drops a part, gets confused, or encounters something unexpected, it usually pauses and waits for a human. Self-recovery is research-active but not solved.
• Stairs and unusual surfaces. Better than five years ago. Still nowhere near a fit adult human.
• Long-horizon tasks. A robot can pick a part. A robot is harder pressed to assemble a complete IKEA bookshelf from box to finished product without intervention.
• Cost. The most-quoted target is “humanoid robot at the price of a used car.” Tesla and Unitree are pushing toward this. Figure, Apptronik, and Boston Dynamics are not yet competing on price.
• Safety certification for human spaces. Factories that deploy humanoids work around carefully-cordoned zones. Hospitals, schools, and homes require ISO-standard safety regimes that don’t yet exist for general-purpose humanoids.

None of this is permanent. All of it is the work of the next several years. Reasonable observers disagree on whether the home-humanoid timeline is 2028, 2032, or 2040. Almost no one credible thinks the answer is “never.”

How big is the humanoid robotics market expected to get?

Forecasts vary wildly, and the more confident a number sounds the less you should trust it. Goldman Sachs has published estimates in the $38B-$150B annual revenue range by 2035. Morgan Stanley has projected scenarios reaching into the trillions for the broader category of physical-AI labor over the longer term. The International Federation of Robotics tracks deployments rather than projections, and its 2024 World Robotics report counted humanoid deployments still in the low thousands of units globally — a base rate that any “millions by 2030” claim has to bridge.

The forecasts that matter most in 2026 are the production-capacity announcements from the companies themselves, because those are commitments backed by capital expenditure. Figure’s BotQ is built for 12,000 robots in year one, scaling to 100,000/year. Tesla has stated a goal of 1 million Optimus units per year at Fremont by late 2026 (treat as ambitious). Hyundai is building a $26 billion U.S. operation that includes a robotics factory targeted at 30,000 robots/year. Unitree is already shipping G1s to research labs globally at $16,000 base price.

Add those up and the production-capacity build-out alone, if hit, puts the industry on a trajectory of several hundred thousand humanoid robots manufactured per year by 2028. Whether the demand-side matches the capacity-side is the unanswered question.

How does China factor into the humanoid race?

China is unambiguously the second pole of humanoid robotics, and on some dimensions the leading one. Chinese companies (Unitree, UBTech, Fourier Intelligence, Galbot, AgiBot, Xiaomi’s CyberOne) are shipping units at price points the U.S. companies cannot match. Unitree’s G1, priced at ~$16,000, is a research-lab-grade humanoid that’s filtered into universities and startups worldwide. Chinese state policy has explicitly identified humanoid robotics as a strategic priority through the “Made in China 2025” and subsequent industrial policy frameworks.

The U.S. advantage as of 2026 is in foundation-model software (the brains) and in enterprise deployment partnerships (BMW, Mercedes, Toyota, GXO, Amazon). The Chinese advantage is in hardware supply-chain integration, actuator cost, and unit-economics ramp. Whether one side maintains its lead, or whether the two ecosystems converge on parity, is the strategic question the field is watching.

What should you watch for in 2026-2027?

• The first independent third-party verification of a humanoid robot at production scale. Figure’s BMW numbers are reported by Figure and BMW; we have not yet seen externally-audited unit-economics data.
• Tesla Optimus V3 reveal and the gap between announced volume and shipped volume. Musk’s track record on timelines is poor; the reveal itself will be informative.
• Whether Apptronik’s Mercedes pilot expands to production scale. Mercedes is the second auto OEM to commit publicly after BMW; the second deployment will indicate whether automotive humanoids are a category, not just a marketing demo.
• The first hospital or healthcare humanoid pilot. Healthcare has been the longstanding “next vertical after factories” promise. The first credible deployment will indicate whether the category breaks out of industrial settings.
• Whether home humanoid pre-orders convert to actual home deliveries. 1X’s NEO has been taking pre-orders. The first home deliveries will calibrate consumer-humanoid timing.
• The cost trajectory. Unitree’s G1 at $16K is the price-floor benchmark. Whether American companies can build a competitively-priced humanoid is an open question.
• Foundation model convergence. Whether Skild Brain, Pi Zero, or another platform becomes the “Android of humanoid robotics” — usable across hardware — versus each company keeping its model proprietary.
• Government deployment. Defense, customs, postal services, and warehousing logistics are all credible early adopters at scale.

Frequently asked questions

When will I see a humanoid robot at my local Costco or Walmart?

Realistically, behind the scenes first. Warehouse logistics for big-box retailers is a credible 2027-2029 deployment target. Customer-facing retail humanoid robots — the actual store floor — is more like 2030+, gated by safety certification, public-acceptance research, and the unit economics of having a robot stock shelves versus a part-time human worker. Don’t expect to see one on your next Costco run.

Will humanoid robots replace human jobs?

Yes, partially, and the direction matters. Humanoid robots will displace the most repetitive, physically tiring, or hazardous parts of manufacturing and warehousing jobs first — the parts that humans already don’t enjoy and that are hard to staff. Whether net employment goes down depends on whether the productivity unlock creates new jobs that didn’t exist before. History (the cotton gin, the spreadsheet, the ATM) suggests displacement is real and the eventual job mix shifts; the transition is rough for the displaced workers, and policy matters enormously.

How much does a humanoid robot cost in 2026?

Public list prices are scarce because most humanoids are sold via enterprise contracts, not catalog purchases. Industry estimates for full-capability commercial humanoids in 2026 are in the $30,000-$150,000 range, varying enormously with capability tier. Unitree’s G1 at ~$16,000 sets the floor for “research-grade humanoid you can actually buy today.” Tesla has stated a long-term aspiration of ~$20,000 for Optimus; whether that holds remains to be seen.

Are humanoid robots safe around people?

In structured industrial settings with documented safety protocols — yes, with the same kind of caveats that apply to any industrial machinery. In unstructured public spaces — not yet certified, and the safety regime doesn’t fully exist. ISO 13482 covers some personal-care robots; ISO 10218 covers industrial robots. A general-purpose humanoid working unsupervised among the public is a category most safety standards weren’t written for.

Why is everyone obsessed with the humanoid shape?

The argument is data-driven and infrastructure-driven. Humans have built virtually every workspace on the planet for the human body — doorways, stair heights, work-counter heights, tool handles, dashboards, control panels. A humanoid robot can operate in those spaces without redesign. And the largest available training data on physical tasks is videos of humans doing them. Both arguments push toward the human form. Critics counter that wheeled or non-humanoid forms are cheaper and more reliable for specific tasks — and they’re right for those specific tasks. The humanoid bet is on generality.

Should I be worried about humanoid robots?

Worried in the science-fiction sense — no. The 2026 reality is that humanoid robots still pause and wait for help when their cable runs out of slack. Worried in the labor-economics sense — that’s a legitimate concern that deserves serious policy thinking, and the displaced-worker question is real. Worried about misuse — yes, like any powerful technology; that’s why government deployments and weapons-system applications deserve scrutiny and regulation.

Sources

• International Federation of Robotics — World Robotics annual report and statistics
• BMW Group — Physical AI deployment in BMW production
• Boston Dynamics — Atlas production and partnership announcements
• Hyundai Motor Group — Hyundai AI robotics strategy announcements
• Figure AI — Figure 02/03 product information and BMW deployment
• Agility Robotics — Digit deployment with GXO and Toyota
• Apptronik — Apollo humanoid and Mercedes partnership
• 1X Technologies — NEO humanoid product information
• Sanctuary AI — Phoenix humanoid product information
• Unitree — G1 humanoid product information
• The Robot Report — Industry news coverage

You may also like

• What is a Humanoid Robot? (Glossary)
• What is Embodied AI? (Glossary)
• What is Robotics? (Glossary)
• What is Optimus (Tesla Robot)? (Glossary)
• AI in Drones in 2026: The Complete Guide
• Computer Vision in Autonomous Vehicles
• What Is a Large Language Model?