{
  "id": "physical-ai-robotics",
  "label": "Physical AI & Robotics",
  "description": "AI controlling physical systems across manufacturing, agriculture, construction, healthcare, and service. The largest domain at 27 practices, with industrial robotics and pick-and-place at good-practice. Most practices cluster at leading-edge — sim-to-real transfer, humanoid robotics, and surgical automation are progressing rapidly. Nearly half the practices are advancing, making this one of the most dynamic domains by momentum.",
  "icon": "🦾",
  "filters": [
    "moving"
  ],
  "hasSummary": true,
  "hasExecSummary": true,
  "practiceCount": 27,
  "evidenceCount": 5241,
  "practices": [
    {
      "slug": "automated-manufacturing-processes",
      "name": "Automated manufacturing processes",
      "tier": "leading-edge",
      "trend": "steady",
      "blockerType": null,
      "description": "AI-controlled robotic systems for welding, painting, finishing, and continuous manufacturing process control. Includes adaptive weld path planning and real-time process parameter adjustment; distinct from assembly robotics which handles discrete object manipulation. Scope covers ML-driven adaptive control and process optimisation; traditional pre-programmed robotic welding, painting, and PLC-based process control are out of scope.",
      "evidenceCount": 192
    },
    {
      "slug": "autonomous-and-additive-construction",
      "name": "Autonomous & additive construction",
      "tier": "leading-edge",
      "trend": "accelerating",
      "blockerType": null,
      "description": "AI-controlled autonomous earthmoving, robotic demolition, hazardous material handling, and 3D printing for construction. Includes GPS-guided grading and large-scale additive manufacturing; distinct from site monitoring which observes rather than performs construction work. Scope covers AI/ML-driven autonomy (e.g. perception-based navigation, adaptive path planning); GPS-guided grading and fixed-program 3D printing without ML are out of scope.",
      "evidenceCount": 188
    },
    {
      "slug": "autonomous-harvesting",
      "name": "Autonomous harvesting",
      "tier": "leading-edge",
      "trend": "steady",
      "blockerType": null,
      "description": "AI-controlled robotic systems that autonomously harvest crops, adapting to ripeness, terrain, and plant variation. Includes soft fruit picking and selective harvesting; distinct from precision spraying which treats plants rather than harvesting them.",
      "evidenceCount": 184
    },
    {
      "slug": "bim-ai-augmentation",
      "name": "BIM AI augmentation",
      "tier": "leading-edge",
      "trend": "steady",
      "blockerType": null,
      "description": "AI enhancement of Building Information Modelling for clash detection, design optimisation, and construction planning. Includes automated clash resolution and cost estimation from BIM models; distinct from facility digital twins which model operational rather than design-phase buildings.",
      "evidenceCount": 186
    },
    {
      "slug": "cleaning-and-maintenance-robots",
      "name": "Cleaning & maintenance robots",
      "tier": "good-practice",
      "trend": "accelerating",
      "blockerType": null,
      "description": "AI-powered robots for commercial floor cleaning, facility maintenance, and domestic vacuuming and mopping. Includes autonomous floor scrubbers and intelligent domestic cleaning with room mapping; distinct from service robots which interact with people rather than performing maintenance. Scope covers AI-driven navigation and adaptive cleaning; fixed-pattern or random-walk robots without ML are out of scope.",
      "evidenceCount": 197
    },
    {
      "slug": "cnc-and-additive-manufacturing-optimisation",
      "name": "CNC & additive manufacturing optimisation",
      "tier": "leading-edge",
      "trend": "steady",
      "blockerType": null,
      "description": "AI optimisation of CNC machining parameters and additive manufacturing processes for quality, speed, and material efficiency. Includes toolpath optimisation and build parameter tuning; distinct from digital twin simulation which models processes rather than optimising machine parameters.",
      "evidenceCount": 201
    },
    {
      "slug": "collaborative-and-articulated-robot-assembly",
      "name": "Collaborative & articulated robot assembly",
      "tier": "good-practice",
      "trend": "steady",
      "blockerType": null,
      "description": "Robots that work alongside humans or autonomously to perform assembly, manipulation, and pick-and-place operations. Includes force-sensitive cobots and high-speed delta robots; distinct from welding and finishing automation which handles specialised manufacturing processes. Scope covers AI/ML-enabled perception, grasp planning, and adaptive manipulation; pre-programmed fixed-trajectory industrial robots are out of scope.",
      "evidenceCount": 213
    },
    {
      "slug": "construction-site-monitoring-and-surveying",
      "name": "Construction site monitoring & surveying",
      "tier": "leading-edge",
      "trend": "steady",
      "blockerType": null,
      "description": "AI-powered construction site monitoring using drones, cameras, and sensors for progress tracking, safety compliance, and site surveying. Includes automated progress photography analysis and safety violation detection; distinct from BIM which models design rather than monitoring construction.",
      "evidenceCount": 144
    },
    {
      "slug": "delivery-robots-indoor-and-last-mile",
      "name": "Delivery robots — indoor & last-mile",
      "tier": "leading-edge",
      "trend": "steady",
      "blockerType": null,
      "description": "Autonomous robots that deliver items within buildings and across last-mile outdoor environments. Includes sidewalk delivery robots and indoor courier bots; distinct from drone delivery which operates in airspace rather than on surfaces. Scope covers ML/AI-driven approaches; prior deterministic or rules-based automation is out of scope.",
      "evidenceCount": 193
    },
    {
      "slug": "digital-twin-simulation-and-optimisation",
      "name": "Digital twin — simulation & optimisation",
      "tier": "leading-edge",
      "trend": "steady",
      "blockerType": null,
      "description": "AI-powered digital twins that simulate manufacturing processes, optimise real-time production, and model facilities and infrastructure. Includes physics-based process simulation and real-time parameter optimisation; distinct from BIM augmentation which targets building design rather than operational simulation.",
      "evidenceCount": 200
    },
    {
      "slug": "environmental-monitoring-forestry-and-grounds-management",
      "name": "Environmental monitoring, forestry & grounds management",
      "tier": "leading-edge",
      "trend": "steady",
      "blockerType": null,
      "description": "AI-driven systems for environmental monitoring, forestry management, and autonomous grounds maintenance. Includes pollution detection, forest inventory assessment, and autonomous mowing; distinct from precision agriculture which targets food crop production.",
      "evidenceCount": 200
    },
    {
      "slug": "healthcare-logistics-and-laboratory-automation",
      "name": "Healthcare logistics & laboratory automation",
      "tier": "leading-edge",
      "trend": "steady",
      "blockerType": null,
      "description": "AI-controlled robots for laboratory sample handling, pharmacy dispensing, and hospital logistics including supply delivery and waste management. Includes automated sample preparation and medication dispensing; distinct from surgical and rehabilitation robotics which directly interact with patients.",
      "evidenceCount": 198
    },
    {
      "slug": "livestock-monitoring-and-welfare-assessment",
      "name": "Livestock monitoring & welfare assessment",
      "tier": "leading-edge",
      "trend": "steady",
      "blockerType": null,
      "description": "AI monitoring of livestock health, behaviour, and welfare using sensors, cameras, and wearable devices. Includes automated lameness detection and feeding pattern analysis; distinct from crop monitoring which targets plants rather than animals.",
      "evidenceCount": 186
    },
    {
      "slug": "parcel-and-freight-handling-automation",
      "name": "Parcel & freight handling automation",
      "tier": "leading-edge",
      "trend": "steady",
      "blockerType": null,
      "description": "AI-controlled systems that sort parcels, load/unload containers, and handle freight across logistics operations. Includes mixed-SKU singulation and automated truck loading; distinct from warehouse robotics which manages internal storage and fulfilment.",
      "evidenceCount": 184
    },
    {
      "slug": "precision-agriculture-monitoring-and-targeted-treatment",
      "name": "Precision agriculture — monitoring & targeted treatment",
      "tier": "good-practice",
      "trend": "steady",
      "blockerType": "market-scope",
      "description": "AI that monitors crop and soil health and drives precision application of water, fertiliser, and treatments using drones, satellites, and ground sensors. Includes NDVI analysis and variable-rate application; distinct from autonomous harvesting which performs physical crop collection.",
      "evidenceCount": 221
    },
    {
      "slug": "predictive-maintenance-remaining-useful-life-estimation",
      "name": "Predictive maintenance — remaining useful life estimation",
      "tier": "leading-edge",
      "trend": "steady",
      "blockerType": null,
      "description": "AI models that estimate the remaining useful life of components and systems to optimise replacement timing. Includes degradation curve modelling and multi-sensor fusion; distinct from condition monitoring which detects current anomalies rather than predicting remaining life. Scope covers ML-based predictive models; traditional statistical reliability methods (e.g. Weibull analysis) without ML are out of scope.",
      "evidenceCount": 205
    },
    {
      "slug": "predictive-maintenance-sensing-and-condition-monitoring",
      "name": "Predictive maintenance — sensing & condition monitoring",
      "tier": "good-practice",
      "trend": "steady",
      "blockerType": null,
      "description": "AI that monitors equipment health through vibration, acoustic, thermal, and visual sensors and alerts on anomalies. Includes multi-sensor fusion and threshold-based alerting; distinct from remaining useful life estimation which predicts future failure rather than detecting current conditions. Scope covers ML-based anomaly detection and sensor fusion; traditional threshold alarms and statistical process control without ML are out of scope.",
      "evidenceCount": 196
    },
    {
      "slug": "quality-inspection-autonomous-rejectpass-decisions",
      "name": "Quality inspection — autonomous reject/pass decisions",
      "tier": "good-practice",
      "trend": "steady",
      "blockerType": null,
      "description": "AI systems that make autonomous accept/reject decisions on manufactured items without human review. Includes confidence-gated auto-rejection and borderline case escalation; distinct from defect detection which flags issues for human review rather than making final decisions.",
      "evidenceCount": 162
    },
    {
      "slug": "quality-inspection-defect-detection-and-measurement",
      "name": "Quality inspection — defect detection & measurement",
      "tier": "good-practice",
      "trend": "steady",
      "blockerType": null,
      "description": "AI-powered visual and dimensional inspection systems that detect defects, measure tolerances, and classify quality issues. Includes surface defect detection and automated dimensional verification; distinct from autonomous reject/pass decisions which act on inspection results rather than performing them.",
      "evidenceCount": 197
    },
    {
      "slug": "rehabilitation-assistive-and-prosthetic-robotics",
      "name": "Rehabilitation, assistive & prosthetic robotics",
      "tier": "leading-edge",
      "trend": "steady",
      "blockerType": null,
      "description": "AI-controlled rehabilitation robots, assistive devices, prosthetics, and exoskeletons that adapt to user movement and intent. Includes neural-interface prosthetics and adaptive gait training; distinct from surgical robotics which operates in clinical rather than rehabilitative settings.",
      "evidenceCount": 195
    },
    {
      "slug": "service-companion-and-food-preparation-robots",
      "name": "Service, companion & food preparation robots",
      "tier": "leading-edge",
      "trend": "steady",
      "blockerType": null,
      "description": "AI-powered robots that serve customers in retail and hospitality, prepare food, and provide companionship and basic care for elderly individuals. Includes restaurant service robots and social companion robots; distinct from healthcare robotics which operates in clinical settings.",
      "evidenceCount": 204
    },
    {
      "slug": "simulation-based-robot-training",
      "name": "Simulation-based robot training",
      "tier": "leading-edge",
      "trend": "steady",
      "blockerType": null,
      "description": "Training robots in simulated environments (sim-to-real transfer) before deploying learned behaviours in the physical world. Includes domain randomisation and physics simulation; distinct from digital twins which model existing processes rather than training new behaviours.",
      "evidenceCount": 194
    },
    {
      "slug": "smart-home-orchestration-and-automation",
      "name": "Smart home orchestration & automation",
      "tier": "bleeding-edge",
      "trend": "steady",
      "blockerType": null,
      "description": "AI that orchestrates smart home devices, learns preferences, and automates home environment management. Includes multi-device coordination and adaptive environment control; distinct from facilities management which targets commercial rather than residential buildings. Scope covers ML-based learning and orchestration; static rule-based automation (e.g. IFTTT triggers, scheduled routines) is out of scope.",
      "evidenceCount": 207
    },
    {
      "slug": "structural-health-monitoring",
      "name": "Structural health monitoring",
      "tier": "leading-edge",
      "trend": "steady",
      "blockerType": null,
      "description": "AI-powered continuous monitoring of structural integrity in buildings, bridges, and infrastructure using sensor networks. Includes strain analysis and deterioration prediction; distinct from drone inspection which captures periodic snapshots rather than continuous monitoring.",
      "evidenceCount": 186
    },
    {
      "slug": "surgical-robotics-semi-autonomous-procedures",
      "name": "Surgical robotics — semi-autonomous procedures",
      "tier": "bleeding-edge",
      "trend": "steady",
      "blockerType": null,
      "description": "Surgical robots that perform portions of procedures autonomously under surgeon supervision. Includes autonomous suturing and tissue manipulation; distinct from surgeon-assisted systems which augment rather than autonomously perform. Scope covers ML/AI-driven approaches; prior deterministic or rules-based automation is out of scope.",
      "evidenceCount": 176
    },
    {
      "slug": "surgical-robotics-surgeon-assisted",
      "name": "Surgical robotics — surgeon-assisted",
      "tier": "good-practice",
      "trend": "steady",
      "blockerType": null,
      "description": "AI-enhanced surgical robots that augment surgeon capabilities with precision guidance, tremor compensation, and visualisation. Includes da Vinci-style teleoperated systems with AI overlay; distinct from semi-autonomous surgery which performs procedure steps independently.",
      "evidenceCount": 219
    },
    {
      "slug": "warehouse-robotics-navigation-storage-and-fulfilment",
      "name": "Warehouse robotics — navigation, storage & fulfilment",
      "tier": "good-practice",
      "trend": "steady",
      "blockerType": "market-scope",
      "description": "Autonomous mobile robots, automated storage systems, and goods-to-person solutions that navigate warehouses and fulfil orders. Includes SLAM-based navigation and intelligent order sequencing; distinct from parcel handling which manages shipping and receiving rather than internal warehouse operations. Scope covers AI-driven navigation (SLAM, ML path-planning) and intelligent order sequencing; fixed-path AGVs and simple grid-following robots are out of scope.",
      "evidenceCount": 213
    }
  ],
  "summary": "## Where AI Stands in Physical AI & Robotics\n\nPhysical AI has split in two. One half is machines that do a single bounded job in a space someone can control, and these are now industrial infrastructure rather than experiments. DHL has passed a billion robot picks coordinated by Locus Robotics' LocusONE software. Brain Corp's autonomy software runs more than 50,000 commercial floor scrubbers. Komatsu has commissioned its 1,000th autonomous haul truck. UPS routes 68.5% of its US parcel volume through automated buildings, at 28% lower cost per piece. Intuitive Surgical has 11,710 da Vinci systems installed, and Deere says See & Spray now sits on roughly a third of North American sprayers. In China, 17.3% of 196,000 arc-welding workstations had fully closed-loop adaptive control by 2025. These figures describe fleets, not pilots.\n\nThe other half is everything that has to cope with the open world. There the gap between demonstration and deployment is closing more slowly than the capital flowing in assumes. The figures on how far adoption has spread are stark. A Kardex survey found only 6% of warehouses highly automated and more than 60% still fully manual. A meta-survey in *Sensors* found that 77% of automotive AI vision pilots never reach full production, and IDC found that 88% of civil-engineering AI pilots never do. What sets this domain apart from AI in knowledge work is that the rare edge cases are physical. A document model that is right 95% of the time is useful; a robot that is right 95% of the time stops a line, damages stock or hurts someone. In a study of Fortune 500 retail and logistics deployments, simulation-trained warehouse policies fell from above 92% success to 67–74% on live floors within two weeks. Safety certification is tied to one specific combination of robot, gripper, payload and task, so it does not carry over the way model weights do. And every site needs its own integration, floor preparation and exception handling. Adoption therefore builds slowly and concentrates in large operators that already have capital, standardised stock and engineering depth.\n\nMomentum is building where the environment can be bounded or the task narrowed. Examples include retrofit autonomy kits for excavators and haul trucks from Gravis, Bedrock and Pronto; orchestration and licensing software that sits above mixed fleets; and drone earthworks surveys, which RICS-regulated surveyors now accept as the measurement basis on large UK infrastructure sites. Others are dairy welfare monitoring, where Merck's SenseHub covers more than two million US cows, and narrow food tasks such as Chef Robotics' 100 million-plus meal servings. Momentum is stalling where tasks are open-ended or evidence of benefit is thin: general-purpose humanoids, selective harvesting of delicate crops in open fields, and autonomous tissue work in surgery. The same goes for companion robots for older people, conversational smart-home control and the economics of pavement delivery. Research keeps producing impressive phantom trials and livestreams, but the revenue keeps coming from machines that do one thing.\n\n## What's New, 2026-08-31 to 2026-09-28\n\nThis four-week window produced more sobering evidence than breakthroughs, and most of it was about money. Boston Dynamics pushed its IPO past 2027. Its robotics unit reportedly lost 528.4bn won in 2025, even with revenue from its Stretch unloading robot. Dexterity, whose Mech robot loads trailers in production for FedEx, UPS, Sagawa and Maersk, has estimated 2025 revenue of about $21M against a $1.65B valuation. Serve Robotics' 2026 guidance now sits at $9–10M after a Q2 gross margin of -271%, with two-thirds of addressable orders blocked by integration gaps in restaurant back offices. Starship left its US college contracts over low utilisation, and Avride took on 21 campuses. Symbotic's 10-Q showed one customer supplying 90.5% of quarterly revenue, and a securities suit over its deployment times survived a partial motion to dismiss. Walmart's 400-store SymMicro rollout is not expected before around 2028. In food robotics, Forward Fooding estimates that 40–50% of companies have died after raising $10.7B between them, and global hospitality robot sales fell 11% in 2024. The capital that did move went to software and retrofit businesses. Buildots raised $130M. CloudNC raised $20mn, with Lockheed Martin's venture arm taking part, and is pursuing FedRAMP authorisation. Brain Corp moved to licensing its BrainOS software across Tennant, Nilfisk and ICE Cobotics scrubbers, and Gravis named Holcim and Taylor Woodrow as customers for its excavator retrofit kit.\n\nThe second thread was the gap between bench results and reliability in deployment. In semi-autonomous surgery, an occupancy-network system completed eight consecutive tumour resections in patient-derived phantoms: all 77 cuts had negative margins, with a mean margin error of 1.61 mm. Another system, MACAW, reached 93% success at 304 fragments an hour across 100 physical debridement trials on the da Vinci Research Kit, beating the π0 and π0.5 baselines. Yet a *Journal of Robotic Surgery* review states that no autonomous procedure has yet been performed in a human, and separate reviews place autonomy in spinal surgery and brachytherapy at the research stage. The same pattern ran elsewhere. A Korean battery maker abandoned AI vision inspection after a month of wrong reject/pass calls. A model predicting remaining useful life across 63 oil wells collapsed from R² 0.919 to 0.005 once whole wells were held out of training. At Stuttgart's Humanoid Robots Summit, practitioners said a single industrial skill takes around six months to teach. On the positive side, A3 counted nearly 18,000 robots worth about $1.2B ordered in North America in the first half of 2026, and Interact Analysis forecast bin picking growing at 36% a year. A 25-study meta-analysis found that sensor-based variable-rate nitrogen cut nitrogen use by 18% and lifted profit by 6% with no loss of yield, though only 15–25% of farmers have adopted it.\n\n## Key Tensions\n\n- **The model is no longer the bottleneck.** A Cognex applications engineer argued this month that AI inspection fails on PLC and MES integration, support ownership and agreed grading standards, not on accuracy. A Korean battery maker illustrated the point by abandoning its vision system after a month of false rejects and missed defects. The pattern repeats in structural monitoring: one cable-stayed bridge logged 63,934 alarms over two and a half years, and only 401 were ever labelled.\n\n- **Hardware losses, software margins.** Boston Dynamics' robotics unit reportedly lost 528.4bn won in 2025 despite Stretch revenue, Serve ran a gross margin of -271% in Q2, and Symbotic depends on one customer for 90.5% of quarterly revenue. The layer that is working sits above the machines. Brain Corp now licenses BrainOS across three scrubber makers at self-reported gross margins above 70%, and DHL's billion picks were coordinated by LocusONE software rather than by any one robot.\n\n- **Narrow machines keep beating general ones.** Practitioners from Renault, Wandercraft and Siemens said teaching a humanoid one industrial skill takes around six months, against the days or weeks an automotive line needs, and Tau Robotics' $30-an-hour humanoid cleaning service turned out to be teleoperated. Meanwhile single-purpose restaurant runners pay back in 6–14 months and floor scrubbers in 9–18. Humanoid makers are themselves becoming a market for narrow machines: Interact Analysis counts their purchases of cobot arms rising from 5,000 to more than 43,000 units.\n\n- **Clearance and adoption run ahead of outcome evidence.** Only 3 of 1,357 cleared AI medical devices were tested on patient outcomes, and US payers pay no premium for robotic over laparoscopic surgery. Upper-limb rehabilitation exoskeletons improved impairment scores without improving function in a new multicentre trial, a BMC Geriatrics meta-analysis found no significant reduction in loneliness from companion systems, and a review of 125 studies found that claimed smart-home energy savings of 15–25% come mostly from simulations. In these areas adoption is driven by labour shortages and demography rather than demonstrated benefit.\n\n- **Retrofit versus rebuild.** Construction is finding retrofit kits the capital-efficient route, as with Gravis's machine-agnostic excavator autonomy. In buildings and warehouses, retrofitting erodes the business case. Consultant John Santagate says the economics of touchless picking favour greenfield sites, a Thai reseller citing the Madrid Metro deployment estimated 27–39-month payback for cleaning robots on Bangkok transit, and Thai practitioners budget ฿100,000–800,000 of hidden retrofit costs per building. Most of the world's operating floor space already exists, which helps explain why automation stays thin outside mega-sites.\n\n## Top 10 Evidence Items\n\n1. **Inbound Logistics: connected-warehouse roundup with Kardex adoption baseline, Gartner forecast and DHL–Locus 1B picks** (news-coverage) — Anchors the split thesis directly: DHL's billion picks sit against Kardex's finding that 60%+ of warehouses remain fully manual. https://www.inboundlogistics.com/articles/building-the-connected-warehouse-2/\n2. **Chinese industry report quantifies harvesting-robot bottlenecks: under 8% strawberry-picker mass production, 68.5% vision accuracy in harsh conditions, 3.7% cooperative uptake** (industry-report) — Third-party Chinese data shows vision accuracy collapsing in field conditions, grounding why open-world tasks lag industrial infrastructure. https://www.hyccpc360.com/industryFreeDetail/30703.html\n3. **Boston Dynamics delays IPO past 2027 as robotics unit keeps losing money; Stretch carries revenue** (news-coverage) — Shows the hardware-loses-money pattern even where a named product line (Stretch) is generating real revenue. https://www.mitrade.com/insights/news/live-news/article-3-2084034-20260914\n4. **Brain Corp CEO on shifting to BrainOS software licensing for partner floor scrubbers** (news-coverage) — Counterpoint to the hardware losses: Brain Corp's shift to licensing software across three scrubber makers is where the margin actually sits. https://www.roboticsintl.com/article/brain-corp-ceo-david-pinn-outlines-software-revenue-model-after-nine-year-build\n5. **Korean battery maker abandons AI vision inspection after one month over wrong reject/pass calls** (news-coverage) — A live abandonment that supports the tension that integration, ownership and grading standards break deployments, not model accuracy. https://www.chosun.com/english/industry-en/2026/09/15/GZ2LVQQ6NZEUJJ3UVHFLIXWBVQ/\n6. **Ladd Strategic Advisors: Walmart's SymMicro store automation is unproven, with the 400-store rollout gated to around 2028** (opinion) — Walmart's own flagship rollout slipping to 2028 shows even the best-capitalised operator can't shortcut the physical-world integration problem. https://laddstrategicadvisors.com/the-automation-choice-every-executive-at-groceryshop-needs-to-see-clearly/\n7. **Teardown.ai deep dive on Gravis Robotics' retrofit excavator autonomy, funding and customers** (industry-report) — Independent confirmation of retrofit economics working where the environment can be bounded, contrasting with warehouse retrofit struggles. https://www.teardown.ai/companies/gravis-robotics\n8. **Humanoid Robots Summit 2026 Stuttgart: Renault, Wandercraft and Siemens on the pilot-to-production gap** (conference-talk) — Named practitioners quantify the narrow-versus-general gap directly: six months to teach one skill against days for a fixed line change. https://humanoid.guide/humanoid-robots-summit-2026-stuttgart-day-one/\n9. **Occupancy network-guided autonomous robotic partial nephrectomy in patient-derived phantoms** (research-paper) — Represents the strongest bench result in surgery while still falling short of the human-deployment threshold the review elsewhere insists on. https://arxiv.org/abs/2609.16186v1\n10. **Meta-analysis of in-season optical sensor-based variable-rate nitrogen in maize (25 studies, 235 comparisons)** (research-paper) — Shows proven, quantified benefit with unusually thin adoption, illustrating that evidence of gain doesn't guarantee uptake. https://link.springer.com/article/10.1007/s11119-026-10447-1?",
  "execSummary": "**The headline:** Robots that do one job in a controlled space now pay their way. Almost everything else in physical AI is still losing money, and this month the balance sheets showed it.\n\n### The Picture\n\nMost companies have not automated their physical work at all. A Kardex survey found only 6% of warehouses highly automated and more than 60% still fully manual. A small group of large operators is pulling ahead with robot fleets: UPS now moves 68.5% of its US parcels through automated buildings at 28% lower cost per piece, and DHL has passed a billion robot picks. Their edge comes from volume, standardized stock and in-house engineering teams, not from better robots. Everyone else is being pitched general-purpose and humanoid machines that practitioners at Renault and Siemens say need around six months to learn a single industrial skill. The real question for you is which narrow, repetitive task in your operation is bounded enough to automate now.\n\n### This Fortnight\n\n- **Boston Dynamics pushed its IPO past 2027 after its robotics unit reportedly lost 528.4 billion won in 2025.** The loss came despite revenue from Stretch, its truck-unloading robot, which DHL and other shippers use. Treat a vendor's financial health as a selection criterion, because a robot supplier that runs out of cash leaves you with unsupported hardware.\n\n- **Serve Robotics cut its 2026 revenue guidance to $9–10 million and said two-thirds of the orders it could serve are blocked by restaurant back-office systems.** The delivery robots themselves work. What fails is the connection to the systems that handle ordering and pickup. If you pilot service or delivery robots, budget for integration with your own systems as well as for the hardware.\n\n- **Warehouse automation maker Symbotic disclosed that a single customer supplied 90.5% of its quarterly revenue.** A shareholder lawsuit over its deployment times also survived a partial motion to dismiss, and Walmart's 400-store rollout is not expected before around 2028. Before signing a multi-year automation contract, ask the vendor how dependent it is on one buyer.\n\n- **The funding that did close this month went to software and retrofit kits rather than new machines.** Construction-tracking firm Buildots raised $130 million. Brain Corp now licenses its autonomy software across three floor-scrubber brands and reports gross margins above 70% on that software. Most of the lasting value is in the software that coordinates mixed fleets, and that is also where buyers have the most room to negotiate.\n\n- **A Korean battery maker abandoned AI visual inspection after one month of rejected good parts and missed defects.** A Cognex applications engineer argued that the model is rarely the problem. The failures come from connecting it to line systems, deciding who supports it, and agreeing on what counts as a defect. Settle those three questions before you fund an inspection pilot.\n\n### Coming Up\n\n- **From December 2026, EU rules make manufacturers strictly liable for faulty automated product decisions, whether or not anyone was at fault.** The rules cover AI systems that pass or reject products on their own. Under the EU AI Act, those systems are also classed as high-risk, with a compliance deadline of December 2027. Have counsel review any unsupervised accept/reject system now, and keep a human-in-the-loop (a person who reviews each AI output before it ships) on borderline cases.\n\n- **US regulators have proposed banning foreign agricultural spray drones that are already approved, including DJI's, with a 180-day window once the rule is published.** DJI dominates the category, with more than 600,000 spray drones in service worldwide. Farming and land-management businesses that rely on them should price in replacement costs and line up alternative suppliers now.\n\n- **Amazon Web Services shuts down its Lookout for Equipment predictive-maintenance service on October 7, 2026.** Existing users, including Toyota and Koch, have to migrate, and the remaining off-the-shelf options are largely Siemens products. If your equipment-failure forecasting runs on this service, confirm your migration plan this month and use the switch to renegotiate terms.\n\n### What's Hard About This\n\n- **Robots trained in simulation lose much of their performance once they reach real floors.** In a study of Fortune 500 retail and logistics deployments, warehouse robots trained in simulation dropped from above 92% success to as low as 67% on live floors within two weeks. Ask vendors for results from a site like yours, not a demo.\n\n- **A robot's safety approval covers one exact setup and lapses when any part of it changes.** If you change the gripper, the load or the task, the safety assessment has to be redone. That is why every site needs its own integration work, and why costs per site stay high for anyone without engineering depth.\n\n- **In healthcare and elder care, adoption is running ahead of proof that the technology improves outcomes.** Only 3 of 1,357 cleared AI medical devices were tested on patient outcomes, and US insurers pay no more for robotic surgery than for conventional keyhole surgery. Build the business case on staffing and labor savings rather than on promised clinical gains.",
  "headline": "Robots that do one job in a controlled space now pay their way. Almost everything else in physical AI is still losing money, and this month the balance sheets showed it.",
  "execSummarySections": [
    {
      "id": "the-picture",
      "title": "The Picture",
      "body": "Most companies have not automated their physical work at all. A Kardex survey found only 6% of warehouses highly automated and more than 60% still fully manual. A small group of large operators is pulling ahead with robot fleets: UPS now moves 68.5% of its US parcels through automated buildings at 28% lower cost per piece, and DHL has passed a billion robot picks. Their edge comes from volume, standardized stock and in-house engineering teams, not from better robots. Everyone else is being pitched general-purpose and humanoid machines that practitioners at Renault and Siemens say need around six months to learn a single industrial skill. The real question for you is which narrow, repetitive task in your operation is bounded enough to automate now."
    },
    {
      "id": "this-fortnight",
      "title": "This Fortnight",
      "body": "- **Boston Dynamics pushed its IPO past 2027 after its robotics unit reportedly lost 528.4 billion won in 2025.** The loss came despite revenue from Stretch, its truck-unloading robot, which DHL and other shippers use. Treat a vendor's financial health as a selection criterion, because a robot supplier that runs out of cash leaves you with unsupported hardware.\n\n- **Serve Robotics cut its 2026 revenue guidance to $9–10 million and said two-thirds of the orders it could serve are blocked by restaurant back-office systems.** The delivery robots themselves work. What fails is the connection to the systems that handle ordering and pickup. If you pilot service or delivery robots, budget for integration with your own systems as well as for the hardware.\n\n- **Warehouse automation maker Symbotic disclosed that a single customer supplied 90.5% of its quarterly revenue.** A shareholder lawsuit over its deployment times also survived a partial motion to dismiss, and Walmart's 400-store rollout is not expected before around 2028. Before signing a multi-year automation contract, ask the vendor how dependent it is on one buyer.\n\n- **The funding that did close this month went to software and retrofit kits rather than new machines.** Construction-tracking firm Buildots raised $130 million. Brain Corp now licenses its autonomy software across three floor-scrubber brands and reports gross margins above 70% on that software. Most of the lasting value is in the software that coordinates mixed fleets, and that is also where buyers have the most room to negotiate.\n\n- **A Korean battery maker abandoned AI visual inspection after one month of rejected good parts and missed defects.** A Cognex applications engineer argued that the model is rarely the problem. The failures come from connecting it to line systems, deciding who supports it, and agreeing on what counts as a defect. Settle those three questions before you fund an inspection pilot."
    },
    {
      "id": "coming-up",
      "title": "Coming Up",
      "body": "- **From December 2026, EU rules make manufacturers strictly liable for faulty automated product decisions, whether or not anyone was at fault.** The rules cover AI systems that pass or reject products on their own. Under the EU AI Act, those systems are also classed as high-risk, with a compliance deadline of December 2027. Have counsel review any unsupervised accept/reject system now, and keep a human-in-the-loop (a person who reviews each AI output before it ships) on borderline cases.\n\n- **US regulators have proposed banning foreign agricultural spray drones that are already approved, including DJI's, with a 180-day window once the rule is published.** DJI dominates the category, with more than 600,000 spray drones in service worldwide. Farming and land-management businesses that rely on them should price in replacement costs and line up alternative suppliers now.\n\n- **Amazon Web Services shuts down its Lookout for Equipment predictive-maintenance service on October 7, 2026.** Existing users, including Toyota and Koch, have to migrate, and the remaining off-the-shelf options are largely Siemens products. If your equipment-failure forecasting runs on this service, confirm your migration plan this month and use the switch to renegotiate terms."
    },
    {
      "id": "whats-hard-about-this",
      "title": "What's Hard About This",
      "body": "- **Robots trained in simulation lose much of their performance once they reach real floors.** In a study of Fortune 500 retail and logistics deployments, warehouse robots trained in simulation dropped from above 92% success to as low as 67% on live floors within two weeks. Ask vendors for results from a site like yours, not a demo.\n\n- **A robot's safety approval covers one exact setup and lapses when any part of it changes.** If you change the gripper, the load or the task, the safety assessment has to be redone. That is why every site needs its own integration work, and why costs per site stay high for anyone without engineering depth.\n\n- **In healthcare and elder care, adoption is running ahead of proof that the technology improves outcomes.** Only 3 of 1,357 cleared AI medical devices were tested on patient outcomes, and US insurers pay no more for robotic surgery than for conventional keyhole surgery. Build the business case on staffing and labor savings rather than on promised clinical gains."
    }
  ],
  "url": "https://www.thestateofplay.ai/domain/physical-ai-robotics",
  "license": "CC BY 4.0",
  "licenseUrl": "https://creativecommons.org/licenses/by/4.0/",
  "generatedAt": "2026-10-01"
}