Robot brains now ship as APIs, but the drop from 95% in the lab to 60% on the floor is still nobody's job

A buyer with 45 humanoid pilots and no payback yet

Fact On August 5, 2026, GXO Logistics chief executive Patrick Kelleher told analysts on the company's second-quarter call that GXO had run 45 humanoid pilots. "We have not achieved ROI on humanoids yet," he said, and put payback a couple of years out. On the same call GXO said it expected to deploy about 20,000 robots of all kinds across its network in 2026.20

Fact Eleven weeks later, on October 26, Agility Robotics and GXO said they would take the Digit humanoid into two more e-commerce fulfillment sites, starting with Digit 4 in early 2027 and moving to Digit 5 once it is commercially available. Neither company disclosed a robot count or a contract value. At GXO's existing site in Flowery Branch, Georgia, Digit has moved more than 100,000 totes, according to The Robot Report.21

Our read A buyer that says it has not been paid back and expands anyway is buying an option on future labor, and it is also telling vendors where the gap sits. The model is no longer the main hold-up. The work that turns a capable policy into a robot a site manager trusts on every shift is data, evaluation, fleet operations and safety, and in each of those few companies have set up shop. That is the open ground this piece maps.

The model layer is turning into a product

Fact In October 2026 Physical Intelligence opened a beta API that lets customers upload data and use hosted fine-tuning and hosted inference. Access runs through a waitlist and there is no public pricing. In July 2026 Google DeepMind put Gemini Robotics-ER 2 into Google AI Studio, while its vision-language-action model and on-device model went only to early partners.23

Fact Deployment at scale is real, but it is concentrated in a few categories. Waymo went from about 50,000 paid rides a week in May 2024 to about 500,000 in March 2026 across 10 cities. Amazon said in June 2025 that it had passed 1 million robots across more than 300 facilities.410

Our read Both scaled examples are operators that own the robot, the site, the software and the data. Waymo runs its own fleet and Amazon runs its own warehouses. Neither had to sell a robot to a third party and prove payback to someone else's finance team, which is the position every humanoid vendor is now in.

Fact Capital has followed. Penn Mutual Asset Management, citing CB Insights, put 2025 robotics funding at a record $40.7B. Citing PitchBook, it put first-quarter 2026 exits in robotics and Physical AI at about $6.2B across 18 deals, more than 2017 through 2025 combined, and cautioned that a single quarter may be an outlier.5

Our read Once a lab sells model access, its growth depends on customers who can bring task data, run robots on real sites and absorb failures. That is a different business from training models, and the labs have so far left most of it to others. The commercial room is in that middle layer, between a hosted model and a robot that works every shift.

Data has become a supply chain

Fact Figure's Index app pays ordinary people to film themselves doing household and work tasks. In August 2026 Figure reported 264,000 downloads in 108 countries, 44,000 weekly active users, more than 16 million videos and $15M paid to the people filming. It also committed more than $1B over the next 12 months to data and compute.6

Fact A month later Figure published what that data bought. In 30 Bay Area homes its robot had never seen, a policy trained from scratch succeeded on 9% of zero-shot trials, while one pretrained on Index video succeeded on 56%. Figure said the evaluation was blind, gave no partial credit and counted any safety intervention as a failure. It did not say how many trials it ran.22

Fact Dyna Robotics pretrained Dyna-2 on more than 1 million hours of first-person human video, with no robot data in pretraining. In the company's own tests, average normalized scores across 14 tasks rose from 20% to 53% across four data tiers, and zero-shot passes against one customer's production standard went from 46% for Dyna-1 to 87%.7

Fact Physical Intelligence reported in December 2025 that the benefit of human video appears only once robot pretraining is diverse enough. On four scenarios that appeared only in human data, it measured roughly a 2x average gain.8

Our read Demand for varied data at volume is plain, but the buyers are a handful of model companies that also run their own collection, and Figure is spending at a scale few suppliers can match. An outside data business needs access the labs cannot easily get, such as specific factories, warehouses or hospitals, or quality tooling they would rather buy than build. Every scaling curve cited here is self-reported, and none has been independently reproduced.

Evaluation still has no referee

Fact In a January 2026 essay, a16z estimated that a policy scoring 95% in the lab may fall to about 60% in the field. Even at 95%, a robot making thousands of attempts a day would fail about 50 times daily, while manufacturing usually expects reliability above 99.9%.1

Fact Google DeepMind published Gemini Robotics 2 numbers that are far from perfect: 68.4% for picking objects from a table, 45.7% from the floor, and 32% to 92% on multi-finger dexterous tasks. Figure's 56% home result above, while more than six times its baseline, still means a failure on nearly every other attempt.322

Fact Open-source tooling is moving toward evaluation and deployment. LeRobot v0.6.0, released by Hugging Face in July 2026, put six new simulation benchmarks under a single lerobot-eval command and added a lerobot-rollout deployment tool that lets a human take over and correct the robot.9

Our read Buyers today compare benchmarks that vendors run on tasks vendors choose, and the most careful published protocol still leaves out trial counts. A neutral measure of interventions per thousand tasks on the buyer's own floor is something model companies have little reason to provide, and something a buyer like GXO, 45 pilots in, would plausibly pay for.

Fleet operations is where margin leaks

Fact Amazon said in 2025 that it expects its DeepFleet model to improve the travel efficiency of its robot fleet by 10%. In February 2026 it added a note to the launch page of its Blue Jay multi-arm system saying it no longer uses Blue Jay in operations.1011

Fact Symbotic, the largest listed warehouse automation company, reported fiscal third-quarter 2026 revenue of $721M, of which $671M came from systems, $37M from operations services and $13M from software, according to its SEC filing. It also reported its first quarterly net income, $55M.12

Fact In February 2026 Agility moved Toyota Motor Manufacturing Canada from a one-year pilot to a commercial robots-as-a-service agreement with seven Digit robots allocated. Agility keeps ownership of the hardware along with maintenance and software upgrades. Locus Robotics said the same month that DHL Supply Chain had passed 1 billion picks on its mobile robots.1323

Fact a16z notes that plant maintenance teams cannot diagnose neural-network failures. Robot companies are staffing the gap themselves: Dyna's job board lists a Senior Software Engineer, Fleet Connectivity and Robot Safety Operators placed at customer sites.114

Our read The pattern is consistent. Hardware and systems carry the revenue, while software and services are thin, and the vendor ends up carrying uptime. Under RaaS every hour of downtime and every remote takeover comes out of the vendor's margin, which makes fleet monitoring, remote assistance and on-site technicians a cost line that grows with each robot shipped. Tools that cut that cost are a gap with paying customers already attached.

Safety and compliance sit in a standards gap

Fact ISO 25785-1, the safety standard for dynamically stable industrial mobile robots, the category that covers legged humanoids, was still a committee draft as of November 1, 2026. Part 2, on application integration, is being drafted separately, and non-industrial settings fall under ISO 13482.16

Fact a16z points out that the existing industrial robot standards, ISO 10218 and ISO/TS 15066, were not written for learned policies whose behavior is not fully specified in advance.1

Fact Boston Dynamics describes Atlas as still in development and training at Hyundai's application center in Georgia, with parts assembly targeted before 2030. 1X's job board lists a Senior Safety Compliance Engineer.1715

Our read Until a standard is published, every humanoid deployment near people is a case-by-case decision for the buyer, its insurer and its safety team. Test evidence, risk assessments and documentation that help a buyer sign off are a near-term service. The window narrows once the standard lands and certification bodies take over that role.

The economics are still illustrative

Fact Agility's investor presentation, filed with the SEC in October 2026 through its merger partner Churchill Capital Corp XI, assumes $8,500 a month per Digit against about $200,000 a year in fully burdened labor, a 2.0x return under RaaS, and a current Digit 4 bill of materials of about $125,000. The filing labels these figures illustrative.18

Fact Two months earlier, the CEO of GXO, whose Georgia site already runs Digit, had said humanoids were not yet paying back. The roboticist Rodney Brooks wrote in January 2026 that most early humanoid companies will quietly shut down or change what they do.2019

Our read The distance between an illustrative 2.0x in a filing and a buyer saying it has not been paid back is itself a market. Whoever can show buyers audited intervention rates, uptime and full cost, teleoperation and on-site staff included, will shape how this category gets procured, and will be the first to know which vendors can price a fleet contract with confidence.