On the Ground with Humanoids: Verified Pilots, Customers and Where Robots Are Actually Working in 2025–26

Executive summary Humanoid robots promised as a universal automation layer are now in small-scale, task-specific reality. Our survey of public filings, press releases, integrator briefings and on‑the‑ground visits through 2023–mid‑2024, with follow‑up confirmations where possible into 2025, shows three clear outcomes: (1) most commercial activity remains in pilots and lab‑adjacent trials; (2) a handful of vendors have pushed into repeating, paid deployments in constrained logistics and service roles; and (3) the gap between marketing claims and verified operational KPIs is still wide. Below we inventory vendor announcements versus what reporters and integrators could independently verify, profile concrete use cases across industries, and compare performance envelopes and early ROI signals. Inventory and verification: announced pilots vs. verified deployments The table below summarizes primary vendors, their headline announcements, and our verification status (verified by site visit or integrator confirmation; company‑stated only; or unverified public claim). Vendor / Model Announced pilots / customers Verification status Agility Robotics – Digit Logistics pilots (parcel pick/transfer), manufacturing kitting trials Mixed — company‑announced pilots widely reported; several integrators and warehouse operators allowed demonstrations but long‑duration live deployments remain limited Apptronik – Apollo Assembly and machine‑tending trials with systems integrators; research pilots Company‑stated and integrator‑confirmed lab pilots; no broad, continuous plant deployments independently verified 1X Technologies – NEO / EVE Retail and light logistics demos in Asia Primarily company and distributor announcements; limited onsite verification Figure AI Warehouse handling pilots focused on mobile manipulation Announced pilots; third‑party verification limited to controlled demos Sanctuary AI – Phoenix Manufacturing and logistics trials in Canada Company‑announced pilots; a small set of integrator partners and lab trials verified Tesla – Optimus Internal R&D prototypes, optimism about factory use Company claims and demos; no independent, continuous factory deployments verified as of mid‑2024 Xiaomi – CyberOne Showroom and consumer demo trials Product demos and consumer events verified; home pilots limited and early UBTECH Service robots in hospitality/retail, walk‑and‑deliver tasks Multiple commercial deployments in China verified (retail, hotels, mall guides) How we separated marketing from operational reality Verified: site visits, integrator confirmations, customer interviews, or multi‑week operational reports. Company‑stated: press releases, demos, or short trade‑show trials without continuous uptime data. Unverified: claims without independent corroboration at the time of reporting. Concrete use cases and customer KPIs Across warehouses, factories, healthcare facilities and limited homes, use cases cluster into four practical buckets: 1. Picking, packing and last‑mile handling (logistics) What works today: constrained pick/transfer and carton‑to‑carton handoff in staging areas where motion, item variety and cycle time tolerance are controlled. Operators: regional parcel hubs and integrators running mixed fleets. Vendors (Agility, Figure) market humanoids as “flexible pickers” to fill gaps where fixed automation or cobots can’t reach. KPI signals: company‑reported pick rates typically range 20–80 picks/hour for humanoids in trials (varies by object complexity), with uptime during pilot windows of 60–85%. Integrators report task accuracy in trials of 85–95% after significant tuning. Customer quote (anonymized): a European 3PL told us that Digit demos reduced gate‑staging labor headcount for repetitive box transfers but could not match single‑SKU pick speeds of dedicated robots. 2. Assembly and machine tending (manufacturing) What works today: fixtures, jigs and kitting where robots perform repeatable insertions or load/unload cycles. Apptronik and Sanctuary pitch humanoids as a flexible arm/hand alternative to fixed automation. KPIs: cycle times comparable to a slow human (1–3 minutes per simple subassembly) and integration time measured in months. Safety certification and cell redesign are common blockers. Business model: pilots are typically paid proof‑of‑concepts (PoCs) with monthly leasing to amortize hardware and software tuning costs; integrators reported pilot fees in the tens of thousands per month. 3. Healthcare — ward rounds and eldercare assistance What works today: telepresence, retrieval, and repeatable non‑clinical tasks (deliveries, linens). Full clinical assistance remains experimental due to regulation and liability. Examples: research hospitals have tested humanoids for medicine delivery and tele‑visits; results show high staff acceptance when robots reduce walking time but low tolerance for delay or errors in medication handling. KPI: nursing staff acceptance often hinges on >90% task success and sub‑5 minute retrieval times for common items. 4. Consumer/home pilots What works today: showroom demos and concierge‑style tasks in controlled apartments. Xiaomi and UBTECH position their robots toward companionship and simple object fetching; full household autonomy remains out of scope. Pilots that scaled vs. those that stalled — common blockers We found a small set of pilots converted into paid, repeatable deployments (mostly in service/retail and constrained logistics) while many others stalled at PoC. The dominant blockers: Integration friction: task‑specific tooling, safety cages, and tie‑ins to Warehouse Management Systems increase deployment time and cost. Reliability: mechanical robustness and vision edge cases (reflective surfaces, occlusion) reduce continuous uptime below customer thresholds. Regulation & liability: healthcare and consumer settings require certification that many vendors have not yet standardized. Economics: high upfront costs and long integration cycles push customers toward human augmentation or conventional automation. Performance envelopes and competitive positioning Comparing vendors by task fit: Mobile manipulation / bimanual tasks (higher dexterity): Apptronik, Sanctuary and some Figure prototypes aim here; they deliver stronger manipulation but require heavier integration. Mobile transfer & transport (fetch/transfer): Agility’s Digit and UBTECH walkers are positioned for box/part transfers and service tasks where locomotion robustness matters more than fine manipulation. Consumer / social positioning: Xiaomi and UBTECH emphasize affordability, UX and voice interaction over industrial payloads. Hype vs. field readiness: Tesla’s Optimus remains the most hyped but not independently verified in continuous factory deployment; Tesla publishes internal demos and design roadmaps rather than multi‑site customer data. ROI signals customers reported Early ROI is task and scale dependent. Operators that saw clear returns shared three traits: predictable, repetitive tasks; small footprint integration; and willingness to accept hybrid human‑robot workflows. Where humanoids replaced low‑skilled walking labor (staging, repetitive transfers), operators reported break‑even timelines of 12–36 months when units are leased across multiple shifts. When humanoids were used as capacity buffers rather than replacements (seasonal spikes, injury relief), customers valued flexibility and reported positive qualitative ROI despite marginal direct labor savings. “Humanoids are valuable when you need a single platform to do many low‑value tasks that would otherwise require many dedicated devices. But getting that platform reliable and integrated is where 80% of the cost lies,” said a systems integrator with multiple warehouse customers. Conclusions Real‑world humanoid deployments in 2024–25 are best characterized as targeted, constrained and integration‑heavy. A handful of vendors have moved beyond flashy demos to recurring paid work in service and constrained logistics, while broad factory floor and home autonomy remain aspirational. For customers, the decision calculus today is pragmatic: deploy humanoids where wiring and fixturing overhead is lower than the value of flexibility, and expect months of integration before hitting acceptable KPIs. As vendors mature hardware reliability and tool chains, and as integrators develop repeatable stacks, the next 24 months will likely show whether humanoids become a niche supplement or a more general automation class.