Physical AI & Robotics — Biweekly Brief
The headline: Factories have learned to trust AI that watches their machines. Almost none trust it to act on its own. That line, not the technology, decides where automation pays.
The Picture
This is the widest field in AI: robots and smart systems in factories, farms, building sites, hospitals, warehouses, operating rooms, and the home. A clear split has opened. Systems that observe — cameras spotting defects, sensors predicting a breakdown, drones measuring site progress — have gone mainstream: a Gartner survey of 200-plus large companies this quarter found two-thirds have moved past trials into full production, cutting costs 18 to 24 percent within 12 to 18 months. Systems that decide and act without a person checking each output are a different story: in a separate survey of 129 manufacturers, 91 percent run none at all. Most organizations sit in the first group and nowhere near the second, and the mid-market is falling further behind — 52 percent of warehouses are now entirely manual, up from 43 percent two years ago.
This Fortnight
- Monitoring crossed into the mainstream. Two-thirds of large manufacturers surveyed have taken predictive maintenance and quality inspection out of trials and into production, with measurable cost reductions inside 18 months. If your equipment monitoring is still a pilot, you are now behind the median, not ahead of it.
- A twenty-year monopoly in surgical robots ended. Johnson & Johnson won US Food and Drug Administration clearance on July 22 for a surgical robot covering ten general-surgery operations — the first genuinely new device category in two decades, breaking Intuitive's da Vinci hold on soft-tissue surgery. Competition should ease pricing, but no insurer pays a premium for robotic over keyhole surgery, so the business case still rests on throughput.
- Washington blocked Chinese robot imports. On July 28 the Federal Communications Commission added Chinese-made humanoid robots to its restricted list, halting new imports from the two largest suppliers while allowing existing units to run until January 2029. Chinese firms ship roughly 85 percent of the world's humanoids and undercut Western rivals by 40 to 60 percent on price, so check where your vendors' hardware is actually manufactured before committing to a multi-year rollout.
- The biggest construction-tech deal ever was about data, not drones. Procore paid $845 million for DroneDeploy, combining 400 million site photos with 20 trillion square feet of labeled aerial imagery. Whichever platform you pilot on ends up owning your project record — negotiate data portability into the first contract.
- Independent scorekeepers arrived, and the numbers were smaller than the press releases. Three public registries launched or expanded this fortnight counting robots verifiably installed and running in construction, surgery, and warehousing — one logging 24 safety incidents alongside 52 surgical deployments. Ask vendors for registry-verified reference sites, not demo videos.
Coming Up
- European rules land in sequence over the next 18 months. The EU Data Act takes effect in September and changes who owns the sensor data your robots collect; strict liability for deployers — meaning you, not the vendor, answer when an autonomous system gets a decision wrong — starts in December 2026; and autonomous quality inspection is now formally high-risk under the EU AI Act, with a December 2027 deadline and fines up to 15 million euros or 3 percent of global turnover. Map now which of your systems make a call without a person reviewing it.
- A vendor shakeout is under way at the frontier. Vicarious Surgical liquidated after roughly $300 million spent and no patients ever treated; food-robot startup Picnic went bankrupt; Ekso Bionics, the largest institutional exoskeleton supplier, exited medical robotics. Check the balance sheet and put service-continuity terms in the contract before signing anything multi-year.
- Healthcare reimbursement is starting to move. US Medicare regulators created a payment category for algorithm-driven clinical care in the 2027 outpatient rules, Japan begins reimbursement pre-screening for certified walking exoskeletons in October, and China's newly approved brain-implant device secured reimbursement within 48 hours. Business cases rejected on economics 18 months ago are worth revisiting.
What's Hard About This
- The robot is the cheap part. Integration typically costs two to five times the hardware. Practitioners now advise budgeting 40 to 60 percent of a project timeline for making a lab-trained robot work on a real floor, against the 10 to 20 percent most plans assume.
- Laboratory accuracy does not survive the shop floor. Robots trained in simulation at Amazon, Walmart, Target, and Ocado dropped from over 92 percent success to 67–74 percent within two weeks of going live. On dusty building sites, vision systems marketed at 95 percent accuracy deliver closer to 80. Insist on a paid trial in your own conditions before scaling.
- Front-line trust is the quiet ceiling. In electronics assembly, the rate at which operators override the machine's quality decision has risen from 12 percent to 41 percent as false alarms accumulate; maintenance technicians routinely switch off alert systems that cry wolf. Systems that lose the confidence of the people using them get bypassed regardless of how accurate they are on paper.
Go deeper: the full Physical AI & Robotics briefing — the longer analytical write-up, plus every practice we track in this domain with its maturity rating, the tools to consider, and the evidence behind our assessment.