From text to bodies: the move to embodied AI

The same trick, pointed at the physical world

The large language model story is familiar by now. Train a big network on internet text and general abilities fall out. Multimodal models extended the trick to images and audio. The next step looks obvious in hindsight: if a model can read a picture and a sentence, maybe it can also write a motor command.

Google DeepMind made that concrete in July 2023 with RT-2, which it called a vision-language-action model. The idea was simple. Represent robot actions as tokens, the way a language model represents words, and train on web data and robot data together. On objects and scenes it had never seen, RT-2 succeeded 62% of the time, against 32% for its predecessor, RT-1.

The field has moved fast since. In October 2024 Physical Intelligence published π0, a model built on a pretrained vision-language model. It was trained on data from single-arm robots, dual-arm robots and mobile manipulators, and showed results on folding laundry, clearing tables and assembling boxes. On March 12, 2025, Google DeepMind announced Gemini Robotics, which adds physical actions as a new output of Gemini 2.0. Six days later NVIDIA released Isaac GR00T N1, which it described as the first open foundation model for humanoid robots.

The direction is clear. Robot control is moving from hand-written code for each task to one general model per robot, or one model across many robots. Software went through the same shift when we stopped writing rules and started training models. The open question is how much of that shift survives contact with the physical world.

The capital arrived before the revenue

Money followed the papers. Figure said in September 2025 that its Series C exceeded $1 billion in committed capital at a $39 billion post-money valuation. Physical Intelligence has raised over $1 billion and was valued at $5.6 billion, according to TechCrunch. Waymo raised $16 billion in February 2026 at a $126 billion post-money valuation. In June 2026 Agility Robotics agreed to go public through a merger with Churchill Capital Corp XI at a $2.5 billion pre-money equity value.

Most of those numbers price a future, not a present. Physical Intelligence co-founder Lachy Groom told TechCrunch he does not give investors answers on commercialization, and the company has no timeline for revenue. I respect the honesty. It also tells you what stage this is. Outside driving, most of the value in embodied AI today is option value on research that has not yet met a paying customer at scale.

Count the robots, not the videos

The most useful habit I took from evaluating more than 150 early-stage startups is to ask for the denominator. A video shows one success. A deployment shows the success rate, the operating hours and what broke.

Figure's own write-up of its BMW pilot is one of the most useful documents in the humanoid industry, because it gives the denominator. A single Figure 02 robot worked at BMW's Spartanburg plant for about ten months on 10-hour shifts, Monday to Friday. It logged more than 1,250 hours, loaded more than 90,000 sheet-metal parts to a 5 millimeter tolerance, and contributed to more than 30,000 BMW X3s. The forearm was the main failure point, and Figure redesigned the wrist electronics for Figure 03 as a result. That is real work in a real plant. It is also one robot doing one task, set against a $39 billion valuation.

Agility Robotics is further along commercially. Its Digit robots moved more than 100,000 totes for the logistics company GXO in 2025. By June 2026 Digit had logged more than 65,000 hours across nine customer facilities, and Agility says it had $300 million in multi-year customer orders as of May 2026. Those are the most concrete humanoid numbers I have seen.

In homes, 1X opened orders for its NEO humanoid in October 2025 at $20,000, or $499 a month, with first US deliveries planned for 2026. The detail that matters is on the product page: for tasks NEO does not know, a 1X Expert can remotely supervise it at scheduled times. Teleoperation as a bridge to autonomy is a sensible way to ship. It also means a human is still in the early unit economics. At Tesla, Elon Musk said on the Q4 2025 earnings call that existing Optimus units were primarily for learning, not productive tasks. Tech Times reported that as of July 20, 2026, Tesla had published no Optimus production count.

The best model teams are candid about limits. When Physical Intelligence released π0.7 in April 2026, one of its researchers told TechCrunch, "You can't tell it, 'Hey, go make me some toast.'" In one air fryer test, success went from 5% to 95% after researchers spent half an hour refining their prompts. That is impressive. It also shows that the person writing the prompt is still part of the system.

Meanwhile the old robot business keeps growing. The International Federation of Robotics counted 542,000 industrial robots installed in 2024 and 4,664,000 in operation worldwide. Most of them are conventional machines running programmed tasks, and they are the installed base that any embodied AI company has to beat on cost and uptime.

Waymo shows the curve can bend

Driving is the one place where embodied AI is clearly a deployed business. In February 2026 Waymo said it was giving more than 400,000 rides a week across six US metro areas, had completed 15 million rides in 2025, and had logged 127 million fully autonomous miles. It also reports a 90% reduction in serious injury crashes. That is the company's own analysis, so I treat it as a claim, but it is a claim others can check. By September 2026 Waymo said it was delivering over half a million trips a week and operating fully autonomously in 15 cities.

The lesson I take for robotics is about the path. It ran through one narrow domain, an enormous amount of real operating data, and years of spending before the numbers worked. Waymo still raised $16 billion in a single round this year. Robot companies that expect to skip that phase are, in my view, mispricing the work.

The five bottlenecks, stated plainly

Data comes first. Language models had the internet. Robots have no internet of actions, so every company is building its own. NVIDIA's answer is simulation: it said it generated 780,000 synthetic trajectories in 11 hours, equal to about 6,500 hours of human demonstrations, and that combining them with real data improved performance by 40%. Physical Intelligence collects data on robot stations in its own office and in warehouses and homes. Figure said part of its Series C would fund human video and other data collection. My view is that synthetic data narrows the gap without closing it, because the hard part of manipulation is contact, and contact is what simulators get least right.

Hardware cost is second. Groom told TechCrunch that the robot arms Physical Intelligence uses sell for about $3,500 and would cost under $1,000 in materials if built in-house. Cheap arms are a gift to research. Humanoids carry far more actuators, and the only consumer humanoid price I can verify is NEO at $20,000.

Reliability is third, and it is the one demos hide. Figure's forearm is the honest version of this problem. Battery life is part of it too: Agility says its new Digit 5 runs for 90 minutes on a 9 minute charge. A factory runs shifts, so the robot has to match the shift, whatever the demo showed.

Safety is fourth. Industrial robots work behind fences, and humanoids are sold on the promise that they will not need them. Agility says Digit is the first humanoid to pass an independent OSHA field evaluation, and it contributes to standards still being written, including ISO 25785-1. Google DeepMind released a dataset called ASIMOV to measure the safety of robot actions. Until those standards settle, every buyer's safety team writes its own rules, which slows every sale.

Unit economics is fifth, and it sums up the other four. A robot earns its place only when its cost per useful hour, including downtime, supervision and integration, beats the labor or fixed automation it replaces. Robots-as-a-service pricing makes the purchase easier for buyers but moves the risk onto the vendor's balance sheet. Remote supervision helps a product ship but puts a wage back into the cost line.

Where I think value settles

My bet is that the general robot model becomes important but is not the whole prize. GR00T N1 is open. Google DeepMind works with outside robot makers on Gemini Robotics, including Apptronik, with Agility and Boston Dynamics among its trusted testers. When the largest platforms give away or license the brain, the brain alone is hard to defend.

I expect value to collect in three places. The first is whoever owns a deployment and the data it produces, because operating hours are the scarce input. The second is the layer every robot needs whichever model wins: hands and actuators, simulation, evaluation and safety certification. The third is companies that pick one job with a clear labor cost, such as moving totes or loading sheet metal, and own it end to end.

The humanoid form earns its cost where buildings were designed for people and cannot be redesigned. Everywhere else, I expect wheels and arms to win on price for a long time.