| Demo Video Claim | Verified Operational Reality (2026) |
|---|---|
| Robot works a full 8-hour shift non-stop | Agility’s Digit runs roughly 4 hours per charge, requiring a rotating fleet to cover a shift |
| One robot equals one full-time worker | Facilities run a 2:1 ratio of working-to-charging units today, with a 4:1 and eventually 10:1 ratio as a future target, not current fact |
| Robot handles any object it sees | Deployments are scoped to narrow, repeatable tasks such as moving totes or bins between fixed points |
| Fully autonomous, no human involved | Many pilots still rely on remote supervision, teleoperated correction, or human oversight for edge cases |
| Ready to buy and deploy today | Most named deployments remain proof-of-concept or early Robots-as-a-Service pilots, not general commercial availability |
Why the Reel and the Warehouse Floor Tell Different Stories
Every humanoid robotics company now produces the same kind of video: smooth lighting, a single robot gliding through a task, upbeat music, and a caption implying imminent, effortless scale. These clips are marketing assets, not operations reports. They are shot after dozens of takes, often with a human operator steering some or all of the motion through teleoperation, and they never show the robot pausing, falling, or being walked back to a charging dock by a technician.
The more useful signal is what companies report to logistics customers, investors, and trade press once a robot has been running in a live facility for months rather than minutes. As of mid-2026, three named deployments give us enough public detail to compare demo-reel claims against operational reality: Agility Robotics’ Digit at GXO Logistics, Amazon, and Toyota; Figure AI’s humanoid at BMW’s Spartanburg plant; and Apptronik’s Apollo at Mercedes-Benz assembly sites and in a GXO proof-of-concept.
This piece is a reality check, not a hype piece. For the broader industry narrative and market-size projections, see our overview of the 2035 humanoid workforce projections. Here, the focus stays narrow: what is actually running today, how well, and how that compares with what gets shown on stage.
The Named Deployments With Public Numbers
Most humanoid robot companies do not publish detailed uptime or error-rate figures. The handful that do give us a rare, if incomplete, window into production reality.
| Company / Robot | Site | Reported Status (2026) | Task |
|---|---|---|---|
| Agility Robotics / Digit | GXO Logistics, Flowery Branch, Georgia | Live Robots-as-a-Service contract; passed 100,000 totes moved | Moving totes between conveyor and pallet locations |
| Agility Robotics / Digit | Toyota facility | Completed a year-long commercial pilot | Intralogistics tote handling |
| Agility Robotics / Digit | Amazon fulfillment sites | Testing under an Amazon-backed investment relationship | Moving empty yellow totes |
| Figure AI | BMW Group, Spartanburg, South Carolina | Active pilot for material handling and parts transfer | Sheet metal and parts movement |
| Apptronik / Apollo | Mercedes-Benz European assembly plants | Working designated task zones bounded by light curtains | Multiple task categories in assembly logistics |
| Apptronik / Apollo | GXO Logistics (proof-of-concept) | Early-stage lab evaluation ahead of distribution-center deployment | Under evaluation, not yet live in a distribution center |
Two things stand out immediately. First, none of these are unsupervised, facility-wide deployments; every one of them is scoped to a specific task, a specific zone, or a specific shift. Second, the most quoted number in the industry, Digit’s 100,000-plus totes moved at GXO, is a cumulative count over an extended commercial run, not a per-shift or per-hour throughput figure, which makes it easy to misread as more dramatic than it is.
Uptime Math: The Number Demos Never Mention
The single most important operational metric in warehouse robotics is not top speed or payload. It is uptime, meaning the share of scheduled time a robot is actually doing productive work rather than charging, faulting, or waiting for a human to intervene.
According to reporting on Agility’s own internal targets, Digit currently operates at roughly a 2:1 ratio, meaning two robots are actively working while a third charges, in order to keep a station continuously staffed. Agility has stated a goal of reaching a 4:1 ratio and eventually 10:1 as battery and charging technology improves. That gap between “2:1 today” and “10:1 someday” is the clearest evidence that current battery life, not software intelligence, is the binding constraint on scaling a fleet.
Figure: The Uptime Pipeline Behind One “Working” Robot
A simplified diagram would show three Digit units cycling through a shift: Unit A actively moving totes, Unit B docked and charging for roughly 45-60 minutes after a 4-hour work stint, and Unit C in transition or fault-recovery. To present a single continuously working robot to a customer or in a demo, a facility must actually operate a small fleet behind the scenes, a detail almost never disclosed in promotional footage.
This is also why comparing a humanoid robot to “one full-time employee” is misleading at today’s maturity level. A human logistics worker clocks an 8-hour shift with two short breaks. A single humanoid unit, at roughly 4 hours of runtime per charge, needs a rotating partner just to match that coverage, which means the effective headcount replaced per robot purchased is well below 1:1 until battery swapping or fast-charging closes the gap.
Task Success Rate: The Metric That Rarely Makes the Press Release
Demo videos show a robot succeeding at a task, by definition, because failed takes are edited out. Verified deployment reporting is more candid, if still limited. Agility Robotics has publicly emphasized that its internal benchmarks for GXO center on repeatable material-handling performance across multiple shifts, explicitly framing “uptime, performance, and reliability” as the metrics that matter for a production-grade deployment, rather than a single successful pick shown once.
Apptronik’s Apollo deployment at Mercedes-Benz offers a useful contrast in caution: robots operate inside zones bounded by light curtains and external sensors, and the robot pauses automatically if a human crosses that boundary. That is a deliberate, safety-first design choice, but it also means Apollo is not yet operating with the free-roaming autonomy that concept videos often suggest. The same applies to GXO’s Apollo evaluation, which as of 2026 remains a laboratory proof-of-concept intended to tune the AI model before any distribution-center deployment, not a live warehouse rollout.
Common mistake
Treating a single viral demo clip as evidence of fleet-wide reliability. A 30-second video proves a task is achievable under ideal, curated conditions at least once; it says nothing about the success rate across the 1,000th repetition, in a cluttered aisle, at hour three of a shift, with a low battery.
What worked
GXO’s phased approach with Digit, starting with a narrow, low-risk task (moving empty totes), running it for an extended commercial period, and only then citing a cumulative volume milestone, is the model other operators are now following. Scoping the task narrowly before scaling it is what allowed the 100,000-tote figure to be a genuine operational result rather than a marketing projection.
Demo Claims vs Verified Reality, Side by Side
| Metric | Typical Demo Video Impression | What 2026 Reporting Actually Shows |
|---|---|---|
| Runtime per charge | Implied all-day operation | Approximately 4 hours for Digit before returning to dock |
| Fleet ratio needed for continuous coverage | Not shown; implied 1 robot = 1 station | Roughly 2 robots working per 1 charging today (2:1), improving toward 4:1 |
| Task scope | General-purpose, “does anything” framing | Narrow, repeatable tasks: totes, bins, parts transfer |
| Human oversight | Rarely shown or mentioned | Remote supervision, safety zones, and light-curtain boundaries common |
| Deployment stage | Framed as commercially available now | Mix of live RaaS contracts (Digit/GXO), active pilots (Figure/BMW), and lab-stage proof-of-concept (Apollo/GXO) |
Why the Gap Exists, and Why It Is Shrinking
The gap between demo and deployment is not deception so much as timeline compression. A promotional video shows the destination; a production floor shows the current step on a multi-year path. Three structural reasons explain why the gap is real and why it persists even at well-funded, credible companies:
- Battery physics has not caught up to marketing ambition. Runtime is capped by the same energy-density ceiling affecting every bipedal platform, a topic we cover in a companion piece later in this article.
- Safety certification lags capability demonstrations. A robot can be shown doing a task in a lab long before insurers, plant safety officers, and OSHA-adjacent standards allow it to do that task unsupervised around humans.
- Software generalization is still narrow. Today’s deployed models are tuned to a specific task and environment; broader cross-task competence remains an active research problem, one we unpack separately below.
- Cumulative vs rate metricsA “100,000 totes moved” headline is a lifetime total, not a per-hour throughput figure. Always ask over what time period a milestone was achieved.
- RaaS contract statusRobots-as-a-Service pricing does not necessarily mean full autonomy; it can mean the vendor owns and maintains hardware while human oversight continues behind the scenes.
- Light-curtain boundariesMany “autonomous” pilots are still spatially fenced by sensors that halt the robot the moment a human crosses a line, a safety measure rarely mentioned in coverage.
- Charge-to-work ratioA 2:1 working-to-charging ratio means buying “one robot” does not deliver one continuous worker; fleet economics require planning for the idle units too.
- Teleoperation as a bridge, not a crutchRemote human correction during pilots is a legitimate strategy for collecting training data, not necessarily evidence a deployment is failing; see our piece on teleoperation data collection.
- Facility-specific tuningA robot validated at one GXO site is not automatically ready for a different warehouse layout without additional simulation and retraining, which is why sim-to-real transfer quality varies by deployment.
What This Means for Operators Evaluating a Pilot
For a logistics or manufacturing operator considering a humanoid pilot in the back half of 2026, the practical takeaway is to ask vendors for rate-based and duration-based numbers, not cumulative milestones. Useful due-diligence questions include: What is the measured uptime ratio across a full week, not a best day? What is the task success rate excluding operator-assisted recoveries? How many technician-hours of maintenance does the fleet require per week? What is the actual charge cycle length and how many spare units are needed to keep one station continuously staffed?
Vendors with genuinely mature deployments, such as Agility at GXO, are increasingly willing to share these figures because the numbers, while more modest than a demo reel, are still commercially credible. Vendors who resist rate-based questions and only offer cumulative totals or curated video are usually earlier in the maturity curve than their marketing suggests.
| Question to Ask a Vendor | Why It Matters |
|---|---|
| What is your fleet’s uptime ratio over a rolling 30 days? | Separates marketing snapshots from sustained performance |
| How many charging or swap cycles per shift? | Determines true fleet size needed for continuous coverage |
| What percentage of tasks require human/teleoperated correction? | Reveals actual autonomy level versus assisted operation |
| Is the site load-bearing safety-fenced or open-floor? | Indicates real proximity to unsupervised human-robot interaction |
| What is mean time between hardware faults? | Predicts maintenance staffing and total cost of ownership |
The Trajectory: Real Progress, Just Slower Than the Reel
None of this is an argument that humanoid warehouse robots are vaporware. Digit’s 100,000-plus totes at GXO, a year-long Toyota pilot, and Figure’s ongoing BMW material-handling pilot are genuine, verifiable operational milestones, not staged footage. The honest framing is that 2026 humanoid warehouse robotics is somewhere between a promising pilot stage and narrow commercial deployment, with the trajectory pointed toward broader scale as uptime ratios improve, battery technology matures, and task scope widens. It is simply not yet at the “one robot replaces one worker for a full shift, anywhere in the building” stage that demo footage implies.
- Uptime ratio
- The proportion of a robot fleet actively performing work versus charging or idle at any given time, commonly expressed as a working-to-charging ratio such as 2:1.
- Robots-as-a-Service (RaaS)
- A commercial model where a customer pays a recurring fee for robot labor while the manufacturer retains ownership, maintenance responsibility, and software updates for the hardware.
- Light curtain
- A sensor-based safety boundary that detects when a person crosses into a robot’s designated work zone, triggering the robot to pause or slow down.
- Proof-of-concept (POC)
- An early-stage evaluation, often conducted in a controlled lab setting, intended to validate feasibility before committing to a live-site deployment.
- Teleoperation
- Real-time remote human control of a robot’s movements, frequently used during pilots both to complete tasks and to generate training data for future autonomy.
Key Takeaways
- Agility Robotics’ Digit has moved over 100,000 totes in a live commercial deployment at GXO Logistics, a genuine milestone, but a cumulative one, not a per-shift throughput figure.
- Digit’s uptime ratio is roughly 2:1 today (two working while one charges), with 4:1 and later 10:1 cited as future targets, not current performance.
- Battery runtime of around 4 hours per charge remains the primary bottleneck limiting continuous single-robot coverage of a work shift.
- Figure AI’s BMW pilot and Apptronik’s Apollo deployments at Mercedes-Benz and GXO remain scoped, safety-fenced pilots rather than open-floor general deployments.
- Demo videos systematically omit charging cycles, failed takes, teleoperated assistance, and safety-zone boundaries.
- Operators evaluating a pilot should request rate-based metrics (uptime, success rate, maintenance hours) rather than accepting cumulative milestones or curated video as proof of readiness.
- The technology is making real, verifiable progress, but 2026 deployments sit at narrow, task-scoped commercial pilots, not the general-purpose, full-shift replacement implied by marketing footage.
FAQs
How many humanoid robots are actually deployed in warehouses in 2026?
Verified named deployments remain small in absolute count, concentrated at a handful of sites including GXO Logistics, an Amazon test program, a completed Toyota pilot, and BMW’s Spartanburg plant. Exact fleet sizes are rarely disclosed publicly, but reporting confirms these are live, ongoing operations rather than one-off demonstrations.
What is Agility Robotics Digit’s real uptime rate?
Public reporting indicates Digit currently runs at roughly a 2:1 working-to-charging ratio, with Agility targeting 4:1 and eventually 10:1 as battery and charging infrastructure improve. These are internal operational targets, not independently audited figures.
How long can a warehouse humanoid robot run before recharging?
Agility has publicly cited approximately 4 hours of runtime per charge for Digit, a figure consistent with broader industry battery constraints discussed in coverage of humanoid robot energy density limits.
Is Figure AI’s BMW deployment fully autonomous?
Figure’s pilot at BMW’s Spartanburg plant focuses on material handling and parts transfer under an active commercial pilot arrangement. Public details on the exact level of autonomy versus supervision are limited, consistent with the industry-wide pattern of narrow, closely monitored pilots rather than fully unsupervised operation.
What does the GXO and Apptronik Apollo partnership actually involve right now?
As of 2026, GXO and Apptronik are conducting an early-stage proof-of-concept in a laboratory setting to fine-tune Apollo’s AI model, with a distribution-center deployment planned to follow rather than already underway.
Why do demo videos look so much more capable than real deployments?
Demo videos are curated marketing assets, typically the best take after many attempts, often supplemented by teleoperation, and edited to remove charging cycles, faults, and safety-boundary pauses that are routine parts of real operations.
What questions should a company ask before piloting a warehouse humanoid robot?
Ask for rolling uptime ratios, task success rates excluding human-assisted recoveries, charge-cycle duration, fleet size needed for continuous coverage, and mean time between hardware faults, rather than accepting cumulative milestones or promotional video as evidence of readiness.
Will humanoid robots replace warehouse workers one-to-one by 2030?
Not at current uptime ratios. Because a single unit needs a charging partner to sustain continuous coverage, effective replacement ratios are below 1:1 today; closing that gap depends on battery and charging advances covered in our analysis of bipedal robot battery and actuator limits.
References
- Robotics and Automation News: “Agility Robotics’ Digit humanoid passes 100,000-tote milestone in live GXO implementation”
- Agility Robotics: “Digit Moves Over 100,000 Totes in Commercial Deployment”
- Origin of Bots: “Agility Robotics’ Digit Moves Beyond Pilots, Now Handling Real Warehouse Work at Amazon, Toyota, and GXO”
- The Robot Report: “GXO Logistics, Apptronik test Apollo humanoid robot for warehouse use”
- Startup Fortune: “Apptronik’s Apollo robot has left the lab and is now working factory shifts at Mercedes-Benz”
- CNBC: “Apptronik raises $520 million at $5 billion valuation for Apollo robot”
- Solid Market Research: “Humanoid Robots Cross the Pilot Threshold: Where Factory Deployment Actually Stands in 2026”
For related coverage in this series, see our explainer on robot foundation models, our opinion piece on battery and actuator limits in bipedal robots, and our data-driven look at the humanoid robot cost curve.
