The Real State of Robotics

>by Roman Tsyupryk

>July 25, 2026

This text is just my thoughts out loud. I'm only a human being trying to analyze current information and imagine what might happen in the future. My thoughts could be completely wrong or might be just "noise" or they could be food for brainstorming about "what if..." scenarios.

Where machines that move actually stand — the mature industry nobody notices, the humanoid gold rush everyone notices, and the AI breakthrough quietly connecting them.

1. The Field Is Older — and Bigger — Than the Hype

Robotics in 2026 is really three fields wearing one name:

Industrial robotics — a mature, profitable, half-century-old industry of arms bolted to factory floors
Autonomous mobility — vehicles, drones, and warehouse fleets that navigate the world
General-purpose robotics — the humanoid and AI-driven wave trying to make one machine do anything

The mistake most coverage makes is treating the third as if it were the whole field. The truth is an inverted pyramid: almost all the robots doing real work today are the boring kind.

The numbers (IFR World Robotics 2025):

Metric	Value
Industrial robots in operation worldwide	~4.66 million (2024, +9% year over year)
New installations per year	~542,000 — double the rate of a decade ago, 4th straight year above 500k
Share of new deployments in Asia	74% (Europe 16%, Americas 9%)
China's share of global installations	54% (~295,000 robots — an all-time record)

Buried in those numbers is a geopolitical shift: for the first time, Chinese manufacturers outsold foreign suppliers inside China (57% domestic market share, up from 47% in 2023 — and by other, less-precisely-sourced accounts, from around a quarter a decade ago). The country that was the world's biggest robot buyer is becoming its biggest robot builder.

2. Why Robotics Is Hard: The Moravec Paradox

The foundational puzzle of the field, articulated by Hans Moravec in the 1980s, still rules everything:

Things that are hard for humans (chess, calculus, protein folding) are easy for machines. Things that are easy for humans (folding a shirt, walking on gravel, picking one strawberry without crushing it) are brutally hard for machines.

The reason: abstract reasoning is evolutionarily new and runs on a thin layer of the brain; sensorimotor skill is hundreds of millions of years old and silently uses most of it. We dramatically underestimate manipulation because we're so good at it we can't feel ourselves doing it.

This is why a $500 chess program is superhuman while a $500,000 robot struggles to unload a dishwasher. Every advance in robotics is, one way or another, an attack on this paradox. Which brings us to what changed.

3. The Breakthrough: Robot Foundation Models

Until roughly 2023, every robot behavior was programmed or narrowly trained: one task, one environment, months of engineering. What changed is the same thing that changed language AI — foundation models, now with bodies.

Vision-Language-Action (VLA) models take in camera images and a natural-language instruction ("put the ripe tomatoes in the bowl") and directly output motor commands — joint velocities, end-effector poses, gripper forces. No hand-written pipeline between perception and action. The lineage:

RT-2 (Google, 2023) — showed that a vision-language model fine-tuned on robot data inherits web-scale common sense: it could move a can toward the picture of Taylor Swift without ever being trained on pop stars.
OpenVLA — the open-source 7B-parameter baseline that put VLA research in every university lab.
π0 (pi-zero) from Physical Intelligence — the current high-water mark for generality. It replaced discrete token prediction with flow matching, generating smooth continuous trajectories suited to contact-rich work (folding laundry, assembling boxes), and demonstrated 10+ distinct manipulation tasks across different robot bodies from a single model.
Gemini Robotics (Google DeepMind) and GR00T (NVIDIA) — the big-lab entries; NVIDIA's play is a full stack: Isaac Sim for training in simulation, GR00T as the model, Jetson Thor as the onboard brain.

Three research unlocks power all of these:

Cross-embodiment learning — training one model on demonstrations from many different robot bodies produces a policy that works better on each body than a specialist model. Diversity is not noise; it is signal.
Video pretraining — models like V-JEPA 2 and GR00T's latent-action pretraining learn physical dynamics from ordinary human video (YouTube, egocentric footage) before ever touching a robot — attacking robotics' deepest bottleneck.
Sim-to-real — physics simulators generate millions of trial hours overnight; domain randomization makes the learned skills survive contact with messy reality.

The deepest bottleneck, honestly stated: data. Language models trained on the internet's trillions of words. There is no "internet of actions" — no vast archive of touch, force, and correction. Every major player is racing to build one, mostly through fleets of teleoperated robots recording human demonstrations. Whoever solves robot data at scale likely wins the decade.

4. The Humanoid Gold Rush

The most visible — and most contested — bet in technology: that the right form factor for a general robot is ours, because the entire built world (doors, stairs, shelves, tools) is shaped for human bodies.

Where the leading players actually are (mid-2026):

Figure AI — furthest along in real Western commercial deployment. Figure's own newsroom confirms a Figure 02 deployment at BMW's Spartanburg plant that logged 1,250+ operating hours across 30,000+ vehicles (as of November 2025); Figure 03 units are deployed there too, but neither Figure's nor BMW's official pages state a specific unit count — the widely repeated "40 units, January 2026" figure could not be verified against either company's own site and should be treated as unconfirmed. The oft-cited ~$25 per robot-operating-hour figure likewise appears only in secondary reporting of CEO Brett Adcock's public remarks, not in either company's official materials — a real, plausible number, but company-sourced hearsay rather than an audited figure. Its Helix VLA model maps vision to motion end-to-end, handling sheet-metal parts it has never seen. Its BotQ factory produces a new robot roughly every 90 minutes.
Tesla Optimus — the boldest industrial commitment and the most cautionary gap between ambition and status. Tesla ended Model S/X production in May 2026 to convert Fremont for Optimus V3 manufacturing (22-degree-of-freedom hands, AI5 chip) — yet Musk admitted on Tesla's Q4 2025 earnings call (Jan 28, 2026) that zero Optimus units were doing materially useful work in Tesla's factories, and none are for sale.
Unitree (China) — a volume leader nobody in the West saw coming: 5,500+ humanoids shipped/sold in 2025 (Unitree's own figures put mass-production above 6,500) at prices starting around $13,500 for the G1 (per Unitree's own store), targeting 10,000–20,000 units in 2026. That is more than all Western firms (Tesla, Figure, Agility) combined — but not more than every competitor: rival Chinese firm AgiBot shipped a comparable or larger volume over the same period and is ranked #1 globally by at least one analyst (Omdia), a reminder that China's humanoid race has multiple serious players, not just Unitree. Roughly 90% of all humanoids sold in 2025 were Chinese.
Others — Agility's Digit (warehouse pilots with GXO), Boston Dynamics' electric Atlas (with Hyundai), 1X's NEO (aimed at homes), Apptronik, Sanctuary, AgiBot, and a dozen well-funded Chinese entrants.

Industry forecasts put 2026 humanoid shipments above 50,000 units — a ~700% jump in one year, per TrendForce; other analysts diverge widely (estimates from ~80,000 to ~90,000+), so treat any single number as one forecast among several, not consensus. The honest caveats: most units shipped are research and demo platforms, not workers; reliability data from real deployments is thin; and the $25/hour Figure–BMW figure, while plausible, is company-reported and covers curated tasks with human oversight. The humanoid industry in 2026 is roughly where the automobile was in 1905 — real products, real customers, and no proof yet of the economics at scale.

5. The Robots Already Working

While humanoids take the headlines, the quiet deployments define the present:

Warehouses. Amazon passed its 1 millionth deployed robot in July 2025 — the largest robot fleet in history — plus arms like Sparrow and Robin picking millions of packages. Warehouse robotics is a solved business, not an experiment.

Roads. Waymo offers fully driverless commercial rides in Phoenix, San Francisco, Los Angeles, Austin, and Atlanta, is targeting 20+ cities and ~1 million rides per week by the end of 2026, and raised a $16 billion round to do it. Autonomous driving — declared dead several times — quietly became a working consumer product in geofenced cities.

Operating rooms. Surgical robots assist in millions of procedures annually. Intuitive Surgical's own investor filings confirm a global da Vinci installed base of ~11,106 systems as of December 2025 (Intuitive does not break this out by country in public filings); widely repeated figures for da Vinci's specific installed base within China (commonly cited as "400+ platforms, ~90% share") could not be confirmed against an official source and should be treated as unverified. What is confirmed: Chinese domestic surgical-robot exports grew ~368% year over year. Note what these machines are: teleoperated precision tools, not autonomous surgeons — a reminder that "robot" often means "better human hands," not "no human."

Fields, mines, and disaster zones. Autonomous tractors and fruit pickers (John Deere, Monarch), drone fleets for inspection and delivery (Zipline's medical network), quadrupeds like Spot patrolling plants and radiation zones — each a niche where "dull, dirty, dangerous" stopped being a slogan and became a purchase order.

6. The Unsolved Problems Nobody Should Skip

Hands. The human hand — 27 degrees of freedom, thousands of touch receptors, self-healing skin — remains unmatched. Dexterity, especially with force feedback (feeling how hard you're squeezing), is the field's hardest hardware problem. Watch a demo: if the robot never manipulates anything soft, deformable, or slippery, the hard part was skipped.

Reliability. A demo needs to work once; a product needs 99.9%+ success across thousands of repetitions. Most public humanoid videos are heavily curated. The gap between "can do it" and "can be trusted to do it unattended, all shift, every shift" is where robotics companies die.

Batteries. Most humanoids run 2–5 hours per charge. A robot that works a third of a shift then docks is a very expensive coffee-break machine.

Cost vs. labor. A humanoid must beat not just human wages but purpose-built automation — a conveyor belt is a very cheap robot. The humanoid bet only pays where flexibility genuinely matters.

Safety and standards. A 60 kg machine swinging steel limbs beside humans has no mature certification regime yet; today's deployments keep robots caged or distant. Home robots raise all of this plus privacy — a robot is a camera with arms.

7. An Honest Assessment

Where the field genuinely is, mid-2026:

Proven and scaled: factory arms, warehouse fleets, robotaxis in favorable cities, surgical teleoperation, drones.
Working, early economics: humanoids in structured factory pilots; VLA models generalizing across tasks in labs and first deployments.
Not yet real: the general-purpose home robot; unattended humanoid labor at scale; robot "common sense" about force, fragility, and consequence.

The parallel to language AI is precise and useful: robotics in 2026 feels like NLP in 2019 — foundation models exist, scaling laws are suspected, capital is flooding in, and the products are one or two breakthroughs (mostly in data and hands) away from the discontinuity. Whether that discontinuity arrives in three years or fifteen is the trillion-dollar question, and anyone who claims certainty in either direction is selling something.

See also: The Second Pair of Hands — a short story imagining what daily human life looks like once general-purpose robots finally work.

Sources

Written as a companion piece to The Second Pair of Hands — the facts first, then the dream.

Part of the Still Becoming series. These two articles are just a small part of a larger Still Becoming series exploring how different technologies could complement each other.

→ Next in the series: The Second Pair of Hands