Where Physical AI Actually Works Today: Real Deployments, Real Payback

Search for “physical AI” and you’ll mostly find demo reels: a humanoid folding a shirt, a robot dog opening a door, a warehouse full of orange totes gliding past on autonomous carts. Impressive footage, but footage isn’t deployment. The question that actually matters for anyone with a budget to defend is narrower and more boring: where is this technology running today, at scale, doing real work, and generating numbers a finance team would accept?

The honest answer, as of mid-2026, is: in three places with hard evidence behind them — warehouse and fulfillment logistics, automotive manufacturing, and row-crop agriculture — plus a fast-growing fourth category of humanoid robots doing narrow, supervised tasks inside those first three. Everywhere else, “physical AI in production” still mostly means pilots, trials, and press releases about intent. This article sticks to what’s documented and verifiable, not what’s demoed, and it’s honest about where the public numbers disagree.

Where is physical AI actually deployed at scale right now?

Three sectors have crossed from pilot to production with numbers big enough to independently verify: Amazon’s fulfillment network has more than 1 million robots working across its warehouses, a threshold the company announced in mid-2025 alongside a new generative AI model, DeepFleet, built to coordinate how those robots move. Agility Robotics’ Digit humanoid has logged over 100,000 totes moved in a live, ongoing deployment at a GXO Logistics distribution center in Flowery Branch, Georgia — not a trial, a running commercial contract. BMW has had Figure AI’s humanoids on a live assembly line at its Spartanburg, South Carolina plant since October 2024, first with the Figure 02 model (which contributed to the production of more than 30,000 vehicles) and now with the newer Figure 03, which BMW is also piloting at its Leipzig plant in Germany. And in agriculture, Carbon Robotics’ LaserWeeder — an AI-guided, laser-based weeding machine — had eliminated more than 10 billion weeds across North America, Europe, and Australia by mid-2024, a figure the company has since updated upward.

None of these are demo-day showcases. They’re production systems with multi-year operating histories, and in Agility’s case, financial disclosures filed as part of a public-market listing that let outside analysts check the company’s own claims against contracted revenue. That’s the bar this article uses throughout: at least two independent sources, ideally including a filing, a trade-press outlet, or an on-the-record company statement that a journalist could and did check.

Warehouse and fulfillment: the largest fleet in the world

Amazon’s robotics program is the clearest case of physical AI moving past the pilot stage, because the scale is now large enough that it shows up in the company’s own operational metrics rather than just marketing copy. The fleet — spanning systems like Hercules, Pegasus, and Proteus, which handle everything from moving shelving units to navigating open floor space alongside people — passed 1 million deployed units in mid-2025, according to reporting from TechCrunch and CNBC that corroborates Amazon’s own announcement. Alongside the milestone, Amazon released DeepFleet, a generative AI model designed specifically to coordinate robot routing across a warehouse floor, reducing the traffic-jam problem that emerges once you have that many autonomous units sharing space.

What makes this deployment instructive isn’t the round number — it’s what the round number implies about the underlying economics. A company doesn’t cross seven figures in deployed units on hope. Each additional robot has to clear a return-on-investment bar set by a company famous for relentless cost discipline, repeated across a fleet the size of a mid-sized city’s population. The routing problem DeepFleet addresses is itself a signal of maturity: at low robot density, the hard problem is getting any autonomy to work at all; at Amazon’s density, the hard problem becomes coordination, congestion, and marginal efficiency gains measured in seconds per pick. That’s a different, more advanced set of problems than most companies evaluating warehouse robotics have to solve yet — but it’s also evidence that the base case (does warehouse automation pay for itself?) was settled years ago inside Amazon’s own operations. For a closer look at how warehouse robotics economics actually break down at the deployment level, see our companion piece on warehouse robot deployment economics.

The humanoid-specific data point in this sector comes from Agility Robotics, whose Digit robot has been working a live, ongoing shift at a GXO Logistics distribution center in Flowery Branch, Georgia. By November 2025, Digit had moved more than 100,000 totes at that single site — described by Robotics & Automation News and Interesting Engineering as the most extensively documented set of commercial humanoid deployment hours of any company in the sector. That distinction matters: plenty of humanoid robot companies have shown a robot moving a box once, on camera, in a controlled setting. Very few have a robot that has moved a box more than a hundred thousand times, tracked and reported, at a customer site that isn’t the manufacturer’s own facility.

Is BMW really using humanoid robots on its production line, or is that PR?

It’s real, and it has run long enough to move to a second-generation robot. BMW Group’s Spartanburg, South Carolina plant put Figure AI’s Figure 02 humanoid to work handling sheet-metal parts on a live assembly line starting in October 2024. According to BMW’s own press office and Figure AI’s official announcement — corroborated by Interesting Engineering’s independent reporting — that deployment contributed to the production of more than 30,000 vehicles over the course of the pilot. That’s not a controlled demonstration cell off to the side of the factory; it’s integration into an actual production line building cars that get sold.

In 2026, BMW moved to the next-generation Figure 03 robot at the same Spartanburg plant, reassigning the humanoid to logistics and parts-sequencing work — picking components from unsorted containers and organizing them into sequencing trolleys for line workers. BMW’s press materials describe plans to expand the fleet size at Spartanburg and to evaluate the same approach at its European plants; a pilot has already started at BMW Plant Leipzig in Germany, marking BMW’s first humanoid deployment in a German production facility. The AI Insider’s coverage of the Figure 03 rollout aligns with BMW’s own statements on both the task reassignment and the geographic expansion.

The task shift from “handling sheet-metal parts” (Figure 02) to “logistics and parts-sequencing” (Figure 03) is worth sitting with, because it tells you something about where humanoid robots are actually earning their keep right now. It isn’t the flashiest job on the line — welding, painting, and precision assembly remain the domain of purpose-built industrial robot arms that have done those jobs reliably for decades. It’s the messier, less structured work: picking an item out of a bin that isn’t perfectly organized, carrying it somewhere, and placing it where a human expects to find it next. That’s exactly the kind of task where general-purpose manipulation and mobility — the core promise of humanoid robots — has a real edge over robots built for one fixed motion. It’s also, not coincidentally, a huge share of the labor cost inside a modern factory that has nothing to do with the glamorous parts of “making a car.”

If you’re evaluating whether a collaborative or purpose-built robot arm fits your own production line — a very different, and generally cheaper and more mature, category than humanoids — the practical tradeoffs are covered in our guide to collaborative robots on the factory floor.

What does a real payback story look like for a robotics company, not just a customer?

This is where Agility Robotics’ 2026 disclosures are unusually valuable, because they’re one of the few places in this space where a company’s growth claims have been checked by securities regulators rather than just repeated by trade press. In June 2026, Agility Robotics agreed to go public via a SPAC merger with Churchill Capital Corp XI, a deal that Business Wire, Nasdaq, and an SEC Form 8-K filing all describe as valuing Agility at a $2.5 billion pre-money equity value, with more than $620 million in expected gross proceeds — including roughly $200 million from a PIPE investment at $10 per share led by Foxconn. The combined company is expected to list under the ticker AGLT once the deal closes, pending shareholder and regulatory approval.

The part of the disclosure that matters more than the valuation headline is the order book. As part of the SPAC filings, Agility reported more than $300 million in multi-year contracted orders for its Digit v5 platform, according to Tech Times and GeekWire’s review of the underlying filings — figures subject to contractual milestones rather than guaranteed revenue, but disclosed in a regulated context rather than a press release. A large share of that figure traces to a single three-year, 1,000-robot contract with a customer Agility hasn’t named publicly. Alongside that contract, Agility disclosed active enterprise deployments at Amazon, GXO, Toyota Motor Manufacturing, and Schaeffler — a customer list that spans warehouse logistics and automotive manufacturing, the same two sectors where the rest of this article’s verified deployments sit.

ERP Today’s analysis of the filings frames this accurately: a $2.5 billion valuation built substantially on a single large contract and a cluster of enterprise pilots is a real business, but it’s also a bet that warehouse humanoids are further along the maturity curve than most of the market currently prices in. That’s a fair characterization of the whole physical-AI-in-warehouses story right now — genuinely running production systems, at a scale that’s still small relative to the total addressable market, with concentration risk in the customer list that any serious buyer or investor should account for.

Agriculture: the quietest real deployment in physical AI

Compared to humanoid robots and Amazon’s headline-grabbing fleet, Carbon Robotics’ LaserWeeder gets a fraction of the media attention — which is exactly why it’s worth including here. It’s a robotic system that uses computer vision to identify weeds in row crops and eliminate them with lasers instead of herbicide, and it has been in commercial use since 2022. By mid-2024, the company’s fleet had eliminated more than 10 billion weeds across farms in North America, Europe, and Australia, a milestone confirmed by Business Wire, Global AgTech Initiative, and Design News. Design News quoted CEO Paul Mikesell’s comparison for scale: matching that output by hand would take 100 people ten years of continuous weeding.

Worth being direct about here: the “weeds eliminated” figure is a moving target, and the sources don’t agree on a single number. The 10 billion figure was reported around June 2024; a February 2025 update tied to the launch of the LaserWeeder G2 reportedly put the cumulative count at 15 billion; and at least one lower-quality aggregator source cites 30 billion alongside a “150+ units deployed” claim that no second independent outlet corroborates. Treat that highest figure as unverified. Similarly, descriptions of how widely the LaserWeeder has spread are inconsistent — the company’s own product page describes “100+ growers,” while a 2025 GeekWire piece describes “hundreds of farms in 15 countries.” It’s not clear whether those are measuring the same thing (unique grower accounts versus total farm sites versus deployed machine units), or simply reflect genuine growth between when each figure was captured. Rather than pick whichever number sounds more impressive, the honest read is that Carbon Robotics has a real, growing, multi-country commercial deployment — and that anyone citing a precise cumulative total should be treated with some skepticism about which snapshot in time they’re quoting.

What the LaserWeeder demonstrates that the warehouse and manufacturing examples don’t is unit economics that don’t depend on a single large enterprise customer. Carbon Robotics raised a $70 million Series D round in October 2024, led by BOND with participation from NVIDIA’s investment arm NVentures, bringing its total funding since founding in 2018 to $157 million — a capital structure built on selling or leasing machines to individual farming operations rather than landing one 1,000-unit contract. For a fuller picture of how robotics is spreading across different crop types and farm sizes, see our deep dive on agricultural robotics in commercial use.

Why don’t the “robot ROI” statistics you see everywhere hold up?

If you’ve researched warehouse or cobot automation, you’ve probably seen numbers like “18-24 month average payback period” or “250%+ ROI” or a specific claim that collaborative robots pay for themselves in exactly 195 days. These figures get repeated constantly across SEO-optimized blogs, buyer’s guides, and LinkedIn posts. We checked them against the same two-independent-source bar used throughout this article, and they don’t clear it. Nearly all of these industry-wide payback statistics trace back to a single origin — vendor marketing content, most commonly blog posts published by cobot manufacturer Universal Robots — that gets copied, rephrased, and re-published by affiliate and SEO content sites without any independent verification of the underlying methodology. No trade press investigation or independent analyst research was found corroborating a universal payback window for warehouse or cobot deployments.

That doesn’t mean automation doesn’t pay off — the deployments profiled above obviously do, or Amazon, BMW, and GXO wouldn’t keep expanding them. It means the specific round numbers circulating as industry-wide benchmarks are marketing claims dressed up as research, and applying a vendor’s self-reported average payback period to your own deployment is a good way to build a business case on a number nobody outside that vendor’s marketing team has ever checked. If a payback figure in a proposal can’t be traced to a named company’s actual, disclosed deployment — the way Agility’s contracted-order figures can be traced to SEC filings — treat it as unverified and ask the vendor for their own site-specific numbers instead.

How should you actually evaluate whether physical AI will pay off for your operation?

Given how unreliable the generic ROI statistics are, the more useful approach is asking questions modeled on what actually distinguishes the verified deployments above from the pilots and demos that never scale.

Is there a named customer and a named site, not just a partnership announcement? Amazon’s DeepFleet, BMW’s Spartanburg line, and GXO’s Flowery Branch facility are all specific, checkable places where specific machines do specific work every day. A press release announcing a “strategic partnership” or a “pilot program launch” is a different category of evidence entirely — it tells you a deal was signed, not that the technology works at the site yet.

Does the task match what the robot form factor is actually good at? BMW’s shift from sheet-metal handling to parts-sequencing with its humanoid fleet is a useful signal: even a company running one of the most visible humanoid deployments in the world moved the robot toward less structured, more general-purpose tasks over time rather than more specialized ones. If a vendor is proposing a humanoid for a job a fixed-base arm has done reliably for twenty years, ask why.

Is there a second, independent source for any number you’re being given? This is the simplest test and the one most buyer’s guides fail. A vendor’s own case study, repeated by three blogs that all cite that same case study, is one source, not four. A number confirmed by trade press, a regulatory filing, or a company that isn’t the one selling you the robot is worth more than ten repetitions of the same marketing claim.

Does the deployment have a multi-year operating history, or is it a recent installation still in its honeymoon period? Digit’s 100,000-tote milestone and the LaserWeeder’s multi-year weed-elimination tallies both matter because they reflect sustained operation, not a successful first week. Early performance numbers on a brand-new installation tend to be the best numbers that deployment will ever produce.

None of this replaces a proper pilot on your own floor with your own SKUs, your own facility layout, and your own labor cost structure — no published case study, however well-verified, substitutes for that. But it does tell you which vendor claims are worth spending pilot budget to test in the first place, and which ones are marketing dressed up as data.

What’s still mostly demo, not deployment?

It’s worth being equally honest about the other side of the ledger. Household and domestic humanoid robots, fully autonomous surgical robotics, general-purpose warehouse picking of arbitrary unstructured items (as opposed to totes and sequencing trolleys), and most construction-site robotics remain, as of mid-2026, in pilot or demonstration phases without the kind of multi-year, multi-site, independently verified production history documented above for warehousing, automotive manufacturing, and row-crop weeding. That’s not a permanent state of affairs — Figure’s move from handling sheet metal to open-ended logistics tasks in under two years shows how quickly a narrow deployment can broaden. But if a claim about physical AI can’t point to a named site, a multi-year track record, or a disclosure that outside parties have checked, the honest classification is “not yet,” not “already here.”

The bottom line

The physical AI use cases that have actually crossed from demonstration into durable production share a common shape: a specific customer, a specific site, a task narrow enough to nail reliably, and a track record measured in years and hundreds of thousands of repetitions rather than a single video clip. Amazon’s million-robot fulfillment fleet, BMW’s humanoid-staffed Spartanburg and Leipzig lines, GXO’s 100,000-tote Digit deployment, and Carbon Robotics’ multi-year, multi-country LaserWeeder fleet all meet that bar — and Agility Robotics’ 2026 public-listing filings are a rare case where a robotics company’s growth story has been checked against regulatory disclosure rather than repeated from a press release. Everywhere else in physical AI, and especially around specific ROI and payback numbers, the honest answer is to ask for the second source before you believe the first one. For the fuller map of where physical AI is generating measurable returns across sectors, the applications and ROI hub tracks new verified deployments as they clear the same bar used here, and the glossary is a good next stop for any of the technical terms — payload, dexterity, teleoperation, fleet orchestration — that come up once you start comparing vendor proposals line by line.