In the Field: 2025–26 Humanoid Pilots Doing Real Work — From Warehouse Floors to Care Homes

Executive summary Between 2025–26 humanoid robots moved beyond showrooms into constrained, repeatable pilots across logistics, factory floors, clinics and assisted-living facilities. The work is still experimental, but four patterns emerged from public disclosures, vendor briefings and conversations with integrators and operators: logistics pilots focus on piece-picking and aisle mobility; factory trials center on repetitive co-bot assembly and inspection; eldercare trials emphasize mobility and assistive fetching; and mobility-centric vendors are being tested for uneven-terrain and inspection tasks. Pilots are measured against concrete KPIs (uptime, picks-per-hour, task-success), but most deployments remain hybrid: humans plus humanoids, not humanoids replacing humans. Who is running pilots and what tasks they do Logistics: Agility Robotics (Digit) and 1X Technologies (NEO/EVE) Agility and 1X positioned humanoids as solutions for labor gaps in piece-pick e-commerce and mixed-SKU warehouses. Pilots reported across multiple 3PLs and regional e-commerce fulfillment centers tested: Primary tasks: single-item picks from shelf, tote-to-conveyor transfers, replenishment runs between zones. Typical KPIs shared: 60–150 picks per hour per robot in constrained, standardized bays; 70–85% task success rate on unrigged SKUs; mean uptime 6–10 hours/day depending on charging cycles and human intervention needs. Integration: vision stacks combining depth cameras and dedicated bin cameras, conveyor sight-lines, and WMS hooks via middleware adapters. Robots were most effective when warehouses installed simple fixtures (grip-friendly shelf lips, pick shelves at human heights). Business models: pilots often ran as RaaS trials—monthly fees covering on-site support—transitioning to CapEx offers conditional on throughput guarantees. Operational lessons: successful pilots restricted SKU geometry and controller-exposed task primitives. Where robots were required to handle unpredictable packaging (soft goods, deformables), success rates and picks/hour collapsed. Human–robot handoffs were managed by designated "handoff benches" or static conveyors rather than ad-hoc exchanges. Factory-floor co-bots: Apptronik (Apollo) and Figure AI Manufacturers ran humanoid cobot pilots adjacent to assembly lines and quality-inspection stations. Targets were repeatable insertions, screw-driving handoffs and visual inspection of finished subassemblies. KPIs: cycle-time parity was the clearest gate—robots needed to achieve ≥90% of a trained operator's throughput to be considered. Typical inspection-throughput matched human inspectors for fixed camera angles; for dynamic inspection, Figure and Apptronik reported ~80–95% defect-detection match when in-line camera systems supported the robot. Integration: tight PLC and conveyor control, OPC-UA and ROS bridges, and deterministic ethernet for motion constraints. Safety integration included speed-and-separation fencing during early cycles, moving toward certified collaborative zones as confidence grew. Safety & incidents: a handful of low-speed collisions and near-miss drops in trials prompted mandatory force-limited end-effectors, redundant state estimation and revised human workstation layouts. Eldercare and assisted living: Sanctuary AI (Phoenix), UBTECH and Xiaomi Care homes and private residences are running small pilots focused on non-medical assistance: fetch-and-deliver, guided walking support, and social engagement. Key findings: Use cases: medication reminders, fetching light objects, fall-detection relay, and companionship rounds. KPIs: uptime expectations are higher (10–16 hours/day), but task complexity reduces successful autonomy—assistive fetch success rates ranged broadly (50–80%) depending on household layout and object variability. Integration and privacy: residential pilots required localized edge perception and strict data residency. Facilities demanded opt-in policies and fall-back human-call flows tied to nurse-station systems. Regulatory hurdles: assisted-care pilots faced facility inspections and heightened insurance scrutiny; many providers required liability waivers and staged deployments restricted to staff-supervised use. Tesla Optimus: internal tests and selective public pilots Tesla continued internal factory testing aimed at replacing repetitive tasks in its own assembly operations while running limited public demonstrations. Reported behaviors emphasized mobility and safe collaboration on high-volume, well-structured operations. Outside observers noted Tesla’s approach remains vertically integrated—deep sensor fusion with factory MES and proprietary control stacks—making third-party system integration slower but cohesive within Tesla facilities. Mobility-heavy deployments: Unitree and Boston Dynamics Unitree’s biped/quad hybrids and Boston Dynamics’ Atlas/Spot systems were chosen for mobility-first tasks: uneven-floor inspections, outdoor deliveries on campuses and search-and-rescue tests. These pilots were judged less on picks/hour and more on mission completion, terrain negotiation and uptime under battery constraints. Comparative KPIs and business approaches Vendor Primary domain Picks/Throughput Uptime Business model Agility Robotics (Digit) Logistics 60–120 picks/hr (constrained) 6–10 hrs/day RaaS → CapEx pilot-to-buy 1X Technologies (NEO/EVE) Logistics, piece-pick pilots 50–130 picks/hr 5–9 hrs/day Subscription + integration fees Apptronik / Figure AI Factory co-bot & inspection Varies; target ≥90% human cycle-time 8–12 hrs/day System sales + software support Sanctuary / UBTECH / Xiaomi Eldercare Not primary metric 10–16 hrs/day Hardware sales, care subscriptions Tesla (Optimus) Internal manufacturing Early-stage; proprietary targets Internal targets for continuous shift ops Internal deployment; unclear commercial intent Operational realities: safety, handoffs and facility work Across pilots the recurring operational disconnects were practical, not technical: facilities had to be modified (shelves re-lipped, floors leveled, cables routed) to make tasks repeatable. Human–robot handoffs succeeded when designers removed uncertainty—fixed handoff zones, standardized containers, and human "catch" operators. Safety incidents were mostly low-severity (stalls, trips, soft contact) but triggered conservative risk postures from operators; as a result, many pilots added redundant safety layers and maintained human interveners during critical phases. Adoption signals and where scale is real Repeatable results are clustering where environments can be constrained: fixed-geometry pick bins, standard assembly sequences or structured care facilities. Warehouses and factories with high SKU standardization or process repeatability are achieving measurable ROI in trials. Care and home deployments are further from scale because variability and privacy considerations increase integration cost. Competitive positioning and near-term outlook Vendors market on a spectrum: logistics players (Agility, 1X) sell throughput and labor-gap mitigation; factory players (Apptronik, Figure) sell cycle-time parity and inspection consistency; care vendors (Sanctuary, UBTECH, Xiaomi) sell mobility and staff augmentation; Tesla emphasizes vertical integration. Short-term commercial models favor RaaS plus service because customers demand performance guarantees without heavy upfront risk; long-term CapEx sales remain the end-goal for high-confidence deployments. Conclusion By 2026 humanoid pilots are producing credible, repeatable wins in tightly constrained pockets of logistics and manufacturing. Broader scale depends less on robot locomotion or manipulation breakthroughs than on pragmatic systems engineering—facility adaptation, deterministic integrations with WMS/ERP/PLC systems, and well-defined human–robot choreography. Vendors that pair robots with rigorous integration services and realistic SLAs are closest to converting pilots into multi-site rollouts; for the rest, pilots will remain expensive experiments until the operational unknowns are tamed.