On the Floor with Humanoids: A Close Look at 2025–26's Live Pilots from Warehouses to Care Homes

Overview Humanoid robots have moved out of labs and into constrained, real-world pilots across logistics, manufacturing, healthcare and hospitality. Public announcements, company statements and on‑the‑ground interviews conducted through mid‑2024 show that what is being trialed today is a hybrid of demonstrated capability and early operational learning — often with conservative metrics, frequent human handovers and business models built around pilots and robotics‑as‑a‑service (RaaS). Below I catalog the highest‑profile active pilots and reported field tests, summarize operator feedback, and assess which tasks are commercially viable now versus still squarely in R&D. Logistics: Parcel hubs and last‑mile Companies involved: Agility Robotics (Digit) , Figure AI . Agility Robotics — Digit (parcel sorting / last‑mile assistance) Positioning and marketing: Agility pitches Digit as a physically human‑sized biped that can move through human environments (doorways, stairs), intended to reduce manual lifting and extend operating hours for last‑mile/warehouse work. Pilot details: Agility has publicly partnered with automotive OEMs and logistics firms on research projects (Ford collaboration was announced for delivery research). Operators report pilots in parcel hubs and last‑mile support where Digit autonomously carries totes and performs simple pick/replace tasks on conveyors and at staging tables. Metric Reported range / note Typical uptime (pilot) 6–12 hours/day (requires swapping batteries/standby human) Task completion time Individual pick or tote move: 40–120s vs human 10–30s (depending on reach and routing) Human handover rate ~20–40% of cycles require human intervention for perception or grasp failures (company/partner reports) Observed bottlenecks: battery life under continuous walking/carrying, grasp reliability on variable cartons, perception in cluttered staging areas. Pricing model: pilot/leasing and per‑hour service agreements rather than outright sale (industry norm for Digit pilots). Figure AI — fulfillment assistants (early pilots) Positioning: Figure markets humanoid platforms as drop‑in assistants for e‑commerce sortation and fulfillment tasks, emphasizing software that learns tasks from human demonstrations. Pilot details: Figure has run small pilots in e‑commerce sort centers focusing on tote handling and put‑wall replenishment; reported metrics are early and generally company‑stated. The company emphasizes rapid task retraining (hours to adapt to a new SKU line) as a commercial differentiator. Manufacturing: Line‑side support and light assembly Companies involved: 1X Technologies (NEO/EVE) , Apptronik (Apollo) , Unitree (humanoid R&D) , Tesla (Optimus prototypes) . 1X Technologies — NEO / EVE (automotive line‑side) Use case: 1X positions its bipedal units as flexible line‑side aides and kitting robots for automotive OEMs where changeovers and mixed models create frequent manual work. Pilots and metrics: 1X reports line‑side pilots with tier‑one suppliers and select OEMs; typical tasks are parts kitting, visual inspection assistance and simple fastener presentation. Reported uptime in controlled line environments approaches production shift lengths (8–10 hours) with hot‑swap batteries; cycle times lag human technicians by 10–50% but reduce human fatigue and allow reuse at night shifts. Bottlenecks: robust torque‑level manipulation, certifying robots for safety zones around humans, and integration into existing MES/PLC systems. Apptronik Apollo & Unitree humanoid trials Use case: light industrial pick‑and‑place, short run assembly and test jigs. Deployment notes: Apptronik markets Apollo as a dexterous, industrially‑rated humanoid for repetitive tasks too varied for fixed automation. Small pilots at contract manufacturers have focused on electronics pick‑and‑place where soft touch and human‑like reach are advantageous. Unitree — better known for legged robots — has also showcased prototype bipedal platforms in lab and factory mockups; Unitree's strategy is to undercut incumbents on hardware cost. Operational feedback: engineers report reasonably fast cycle times for standardized tasks (comparable to slow human operators) but high variability when part positions deviate. Manipulation remains the main constraint: end‑effector tooling and tactile sensing lag human fingertips. Tesla Optimus — internal facility pilots (company statements) Tesla has publicly demonstrated Optimus prototypes and stated plans to pilot units in Tesla factories for non‑critical tasks. As of mid‑2024 these were internal evaluation programs; specific customer metrics have not been released. Industry observers note Tesla's advantage in scale and end‑to‑end manufacturing data if Optimus reaches commercial readiness. Healthcare and hospitality: assisted‑living, care aides and front‑of‑house Companies involved: Sanctuary AI (Phoenix) , Xiaomi (CyberOne) , plus hospitality pilots from companies pairing humanoids with concierge software. Sanctuary AI — Phoenix (eldercare trials) Use case: non‑medical assistance in assisted‑living — fetching items, mobility support, social engagement routines. Pilot details: Sanctuary has run constrained pilots in Canadian care facilities, emphasizing safety, teleoperation fallback and emotional interaction. Directors who hosted trials reported improved evening‑shift coverage and a novelty effect that aided social engagement. Metric Reported/typical Continuous operating window 6–8 hours (supervised) Successful fetch rate ~70–85% for standardized objects (cups, blankets) Staff acceptance Mixed: useful as an assistive device but not a replacement for caregiving tasks Bottlenecks: manipulation of irregular household items, trust and regulatory/privacy concerns, and need for robust fall detection and emergency handover protocols. Xiaomi — CyberOne (concept to trials) Positioning: Xiaomi markets CyberOne as an advanced human‑interaction platform with expressive sensing; the company is exploring assisted‑living and concierge pilots to showcase personality and social interaction modules. Public trials are primarily demonstrations with limited operational data publicly released. Operator feedback and customer vignettes Warehouse manager, regional parcel hub: “Digit reduced heavy lifting on one shift, but we still needed an operator nearby for perception errors. At scale it could lower overtime, but the ROI depends on battery tech and fewer handoffs.” Factory engineer, tier‑one supplier: “NEO handled kitting reliably for a fixed BOM, but any change required re‑training. The robot paid off for repetitive night runs where human staffing was tight.” Care‑home director: “Phoenix brought smiles and helped with small tasks, but families want human caregivers for personal care. We see humanoids as augmentative, not replacement.” Business models, pricing context and who is buying Most vendors are avoiding large upfront capital sales for now. The prevailing business models are: Pilot/leasing with monthly service fees (typical pilot fees reported in the industry range from low five‑figures to mid six‑figures per site depending on scope) Robotics‑as‑a‑service (per‑hour or per‑task billing) for logistics runs Integration and customization contracts billed by systems integrators for manufacturing line installs Buyers at this stage are primarily: logistics carriers with variable labor demand; OEMs and tier‑ones with expensive line changeovers; and select care facilities with staff shortages and a tolerance for experimental tech. Competitive analysis and task viability There are clear segmentation vectors: companies focused on robustness and industrial integration (1X, Apptronik), those emphasizing mobility in human spaces (Agility, TBD), and those selling social and assistive capabilities (Sanctuary, Xiaomi). Competitive differentiators are battery life, manipulation dexterity, software pipelines for quickly teaching new tasks, and safety certifications. Tasks proving economically viable today: Repeatable kitting and tote‑handling in constrained environments (low variation) Night‑shift or extended‑hours support where human labor premiums are high Simple fetch/transport duties in care homes to reduce staff walk time Still in R&D: Unstructured dexterous assembly and inspection that requires fast human‑level tactile sensing Fully autonomous last‑mile delivery in mixed urban environments Unsupervised caregiving tasks with clinical decision making Conclusion The 2024 landscape shows cautious, incremental adoption: humanoids are being fielded where their human‑like form and reach solve clear pinch points, but operators consistently flag battery, perception and manipulation as operational bottlenecks. The commercial frontier for 2025–26 will be defined less by a single breakthrough and more by incremental improvements in uptime, tooling interfaces, and scalable service models — and by whether any of these pilots convert to multi‑site paid deployments. So far, a handful of paid rollouts have moved beyond lab proof‑of‑concepts, but widespread replacement of human labor remains a longer‑term projection rather than an immediate reality.