On the Floor, In the Aisles, By the Bed: A 2025–26 Field Report on Live Humanoid Deployments

Executive summary Between 2025 and mid‑2026 the humanoid robotics sector moved from headline demos to a patchwork of customer pilots and limited commercial rollouts. Vendors positioned their platforms as a response to chronic labor shortages, ergonomics and throughput gaps in warehousing and light manufacturing, and staffing pressures in eldercare. On‑the‑ground evidence — vendor announcements, public demos, procurement notices and customer statements — shows pilots remain small, highly supervised and integration‑heavy. Early ROI signals are mixed: productivity gains are visible in narrow, well‑instrumented tasks; total cost of ownership and safety staffing keep most deployments at the “pilot” stage. Why industry is testing humanoids now Labor scarcity and turnover: fulfillment and care sectors with high seasonal peaks and persistent hiring difficulties are prime candidates. Ergonomics and retention: humanoids are marketed as substitutes for repetitive, injury‑prone duties (reaching, bending, lifting small items). Flexible automation: unlike fixed arms, humanoids promise multi‑tasking and human‑centric environments without retooling. Warehousing and logistics Major activity in 2025–26 centered on order‑tending and piece‑picking pilots. Vendors emphasize adaptability to mixed‑SKU aisles rather than outright speed records. Agility Robotics — Digit Agility framed Digit as a "person‑assistant" for inbound/outbound workflows. Public demos and supplier filings show pilots deployed as an order‑tending aid in high‑bay fulfillment centers. Typical assignment: retrieve light cartons from tote dumps, present units to human pickers or conveyor interfaces. Scope: single‑cell pilots (1–3 robots) alongside human pick teams. Performance signals: vendors reported cycle times competitive with novice human pickers on simple reaches; customers emphasised improved ergonomics rather than throughput alone. Integration: moderate — Digit required tethered mapping, conveyor handoff protocols and middleware to interface with WMS; most pilots used manual human verification. Safety & incidents: public records show no major injuries; incidents were limited to navigation stalls and knocked‑over totes handled by supervisors. 1X Technologies — NEO/EVE (order-tending/piece picking) 1X positioned NEO/EVE as a piece‑picking assistant that combines tactile fingers with compliant locomotion. Procurement filings (where made public) and demo footage indicate use in retail replenishment and light e‑commerce picking. Scope: multi‑unit pilots reported in the low single digits; emphasis on selective SKU families. Operational reality: reliable on structured trays; struggled with loose, varied SKUs — human handoff remained common. ROI signals: early customers cited reduced picker fatigue and secondary labor redeployment, but not yet unit‑level payback within 12 months. Discrete manufacturing — kitting and machine tending Manufacturers trialed humanoids for flexible kitting and tending where human‑like reach and two‑hand dexterity matter. Figure AI and Apptronik — factory kitting and tending Both companies marketed humanoids as drop‑in partners for intermittent tasks on factory floors. Scope: factory pilots typically 1–5 robots handling kitting, small part insertion and light machine tending on scheduled shifts. Cycle & reliability: vendors showcased cycle times near human averages for repeatable kitting; real shop‑floor trials reported downtime driven by perception edge cases and tooling misalignment. Integration: higher than vendor marketing suggested — PLC interlocks, safety cages or collaborative zones, and custom end‑effectors were necessary. IT teams spent weeks on data flows and scheduling hooks. Customer feedback: plant managers praised flexibility but warned of ongoing supervision and maintenance overhead. Factory-assist and public demos — Tesla Optimus Tesla’s public demonstrations of Optimus in factory settings attracted heavy attention. Tesla positioned Optimus as an eventual general‑purpose factory assistant; public footage and company statements through 2025–26 indicated limited, supervised factory‑assist pilots. Scope: small numbers of prototype units deployed for noncritical assistance tasks (parts staging, basic transport) rather than autonomous cycle replacement. Operational reality: demos highlighted potential; however, independent observers noted polished scripted sequences versus continuous unsupervised operation. Integration & safety: Tesla reportedly relied on Tesla’s internal networks and close human oversight during pilots; no public record of injury during the period reviewed. Healthcare and eldercare Care homes and assisted‑living sites piloted humanoids for social engagement, mobility assistance and medication delivery under strict clinical oversight. Sanctuary AI and UBTECH — assisted living trials Sanctuary and UBTECH targeted eldercare with different positioning: Sanctuary emphasizing human‑level interaction for complex tasks; UBTECH offering companion and logistics services. Scope: trials often single‑unit in a residential wing or care home, supervised by nursing staff. Workflow integration: significant — clinical workflows required logging, human signoffs, and fallback procedures. Vendors integrated with care management platforms only in limited pilots. Safety & incidents: public disclosures showed conservative operation, with robots largely providing non‑clinical assistance (fetching items, reminders). No major safety incidents were reported in open filings; however facilities reported increased staffing needs for oversight during early shifts. Customer feedback: positive around resident engagement and staff relief on low‑value tasks; skepticism remained about clinical dependence and liability. Consumer-facing and research pilots Companies continued to explore in‑home assistance and research programs rather than commercial rollouts. Xiaomi — CyberOne and research programs Xiaomi framed CyberOne as a research and consumer roadmap device. Reported activity through 2026 focused on developer programs and university partnerships rather than mass consumer trials. Xiaomi’s marketing emphasizes lifestyle and smart‑home integration; practical deployments remained experimental. Boston Dynamics — Atlas integrations for construction/prototyping Boston Dynamics retained Atlas for research integrations with construction and prototyping teams. These were collaborative research projects with universities and engineering firms rather than customer‑facing production pilots. Atlas remained a tool for R&D rather than a deployed site worker. Comparative snapshot Vendor Primary use Pilot scale (typical) Integration effort Reported ROI signal Agility (Digit) Order‑tending 1–3 units Moderate (WMS + conveyors) Ergonomics & redeployment of labor 1X Tech (NEO/EVE) Piece‑picking 1–5 units High for loose SKUs Mixed — works on constrained SKUs Figure / Apptronik Kitting, tending 1–5 units High (PLCs, tooling) Improved flexibility; maintenance costs notable Tesla (Optimus) Factory assist (demo) Prototype clusters High; closed‑loop internal systems Theoretical; scripted demos predominant Sanctuary / UBTECH Eldercare assistance Single‑unit pilots Very high (clinical workflows) Behavioral benefits; clinical ROI unproven Xiaomi / Boston Dynamics Research / demo Lab/partner pilots Very high Long‑term roadmap value Lessons from the floor Across sectors the same themes recur: pilots are small, integration costs dominate, human oversight is non‑negotiable, and safety architectures (geofencing, emergency stops, supervised handoffs) require staffing and training. Where humanoids succeed early is in narrow tasks where human form factor is essential and where the environment is controlled (trayized SKUs, fixed kitting stations, scheduled care rounds). Scaling to production will require: robust perception in clutter, plug‑and‑play integration with enterprise systems, standardized safety certifications, and a services ecosystem for maintenance. Absent those, deployments will remain multi‑year pilots rather than frontline replacements. Outlook For buyers, the pragmatic path in 2026 is staged: pilot a single cell, measure ergonomic and throughput gains, budget for integration and oversight, and treat humanoids as a complement to human teams rather than a turnkey substitute. Vendors continue to close capability gaps; the question for procurement teams is not if humanoids will be useful, but when their total cost profile aligns with predictable, repeatable value — a milestone likely several years away for broad production deployments, but achievable sooner in tightly constrained use cases.