A Figure 03 robot now works a paid shift on BMW’s largest assembly plant in the world, billed at roughly $25 per robot-hour. On Amazon, you can buy a Unitree G1 humanoid for $17,990, delivered like a large appliance. Toyota’s Canadian plant has more than seven Agility Digit robots doing real material-handling work after a successful pilot. None of that was true two years ago. This is what people mean when they say “physical AI” has arrived — and it’s worth a plain-language explanation of what it actually is, why it’s suddenly treated as such a big deal, and how far the technology has really come versus how far the demo reels make it look.
What “physical AI” actually means
Regular AI — a chatbot, an image generator — only has to be right inside a screen. Physical AI has to be right inside a body: it takes in sensor data (cameras, touch, balance) instead of just text, and its output is a motor command instead of a sentence. The industry describes the underlying models as evolving through stages — from language models (LLMs), to models that also understand images and video (LMMs), to “large action models” (LAMs) that output physical actions in the real world. A robot arm that can only weld one specific part on one specific car is old automation. A robot that can look at a bin of unfamiliar parts, a request in plain English, and figure out what to do — that’s physical AI.
Why it’s being treated as urgent right now
The case for physical AI is mostly demographic, and the numbers are blunt. The US alone has roughly 600,000 unfilled manufacturing jobs today, and its eldercare worker shortage is projected to hit 1 million by 2030. Korea, Japan, and China are all aging faster than their workforces are being replenished. Goldman Sachs estimates that even at current robot prices, humanoids could fill about 4% of the US manufacturing labor gap and 2% of global eldercare demand by 2030 — modest on its own, but the direction of travel matters more than the exact percentage.
The other reason is price. A humanoid robot’s average selling price was around $114,700 in 2024; forecasts put it near $37,000 by 2030, a drop of more than two-thirds in six years. That’s not a hypothetical — it’s already visible at the low end: a Unitree G1 runs $13,500–$17,990 today, and 1X sells its home robot NEO for $20,000 outright or $499 a month as a subscription. When a capable robot costs about as much as a used car, the addressable market stops being “big factories” and starts being “everyone’s house,” which is why analysts are throwing out numbers like a $38 billion humanoid market by 2035 (Goldman Sachs) and a $5 trillion opportunity by 2050 (Morgan Stanley).
How good the technology actually is in 2026
The hard problem in physical AI has never been building a robot body — it’s teaching that body to handle situations it wasn’t explicitly programmed for, using far less real-world data than language models get to train on. Nvidia’s Isaac GR00T family of “robot foundation models” is the clearest public window into how the industry is attacking that problem. GR00T N1 uses a dual-system design modeled loosely on human cognition: a slow “System 2” that reasons about a scene and plans what to do, and a fast “System 1” that translates that plan into precise, real-time motor movements. To get around the shortage of real training data, Nvidia generated 780,000 synthetic robot-motion sequences — the equivalent of 6,500 hours of human demonstration — in eleven hours of simulation, and mixing that synthetic data with real recordings improved the model’s performance by roughly 40%. The newest version, GR00T N1.7, is pretrained on 20,000 hours of first-person human video on top of robot demonstrations, and Nvidia has already previewed a next-generation model, GR00T N2, claiming more than double the task-success rate of prior vision-language-action models — expected to ship by the end of 2026.
That foundation-model progress is what’s letting real deployments move past scripted factory demos. Figure’s fleet at BMW and Agility’s robots at Toyota are doing genuine, if narrow, warehouse and assembly work today, billed by the hour like contract labor. Boston Dynamics is scaling up electric Atlas units for Hyundai and DeepMind. Tesla is aiming for volume production of Optimus Gen 3 at its Fremont plant by the end of summer 2026, though it remains deployed internally rather than sold to outside customers. The honest summary: today’s robots are reliable enough for narrow, well-defined, monitored tasks in controlled environments, and still far from the “do anything a person can do” generalist that the marketing implies.
Who’s actually competing, and on what
| Company | Backing / Valuation | 2026 status |
|---|---|---|
| Figure AI | $39B valuation, OpenAI-led $675M round | 40-unit fleet in paid deployment at BMW |
| 1X Technologies | ~$10B, backed by OpenAI, Tiger Global, Samsung | NEO home robot, first US deliveries late 2026 |
| Unitree (China) | Privately held | 5,500+ units shipped in 2025 alone; G1 sold on Amazon |
| Tesla | Public (TSLA) | Optimus Gen 3, internal use only, volume production targeted late summer 2026 |
| Agility Robotics | Privately held | Digit robots doing paid work at Toyota Canada |
| Boston Dynamics | Owned by Hyundai | Electric Atlas fleets for Hyundai and Google DeepMind |
The competitive split that matters most isn’t company-by-company — it’s geographic. The United States holds the clear lead in the “brain”: foundation models from Nvidia, OpenAI, and Google DeepMind, and the multimodal reasoning that lets a robot understand an instruction it’s never seen before. China has taken a commanding lead in the “body”: Unitree alone shipped roughly 35 times more units in 2025 than Tesla did, and Chinese manufacturers are treating 2025-2026 as the first real commercialization wave of humanoid hardware, built on an already-dominant supply chain for motors, actuators, and rare-earth magnets. Betting on physical AI increasingly means picking a side of that split, or a company that can straddle both.
Korea’s position is worth a specific note. It already has the world’s highest industrial robot density — 1,012 robots per 10,000 manufacturing workers as of 2024 — and several analysts now frame it as a plausible “third path” in physical AI, neither the US model (foundation-model-first, hardware second) nor China’s (hardware-and-volume-first), but a manufacturing-integration play built on a robotics base few other countries can match.
What would actually change the picture from here
Watch three things rather than the demo videos: whether Nvidia’s GR00T N2 delivers its claimed jump in task success once it ships, since that’s a direct read on whether the sim-to-real data problem is actually being solved rather than just marketed; whether deployments like Figure’s at BMW or Digit’s at Toyota expand meaningfully in unit count over the next year, since that’s the difference between a pilot and a real product; and whether humanoid prices keep falling on the schedule analysts expect, since the whole “robot in every home” thesis depends on the price curve holding, not just the capability curve.
This article is for informational purposes and is not financial or investment advice. Company valuations, shipment figures, and pricing are drawn from public reporting current as of publication and change quickly in this industry; verify current figures before acting on them.
Sources
- LumiChats — Humanoid Robots 2026: Tesla Optimus vs Figure AI vs Unitree
- ValueAdd VC — Humanoid Robots 2026: Figure vs Apptronik vs 1X vs Tesla Optimus vs Unitree
- NVIDIA Newsroom — NVIDIA Isaac GR00T N1: the World’s First Open Humanoid Robot Foundation Model
- NVIDIA Developer Blog — Develop Humanoid Robot Policies End-to-End with NVIDIA Isaac GR00T
- Tech Times / IDTechEx — Humanoid Robot Price Falls 68% by 2030
- 헤럴드경제 — “2035년 차보다 로봇이 많다”…피지컬 AI 시대, 미·중 대안 부상하는 한국
- 디지털데일리 — [AI 클로즈업] 사람 손 부족한 곳에 피지컬 AI 파고든다