🦾 Physical AI & Robotics
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.
The Headline
Money is reaching robots faster than proof that they pay. A $600 million excavator round and new warehouse robots arrived just as buyers reported cost overruns and field results kept falling short of lab claims.
The Picture
Robots now work reliably where floors are structured, tasks are narrow and a person is close by. Locus says it runs more than 22,000 robots across 360 fulfillment centers. Beyond that core, most organizations are still manual: only about 6% of North American warehouses are highly automated, a share that has not moved meaningfully since 2024. The operators pulling ahead are the largest ones, because they can absorb integrations that average 12 to 18 months and spread the fixed costs. Everyone else faces a market where funding and product launches are moving faster than the evidence that deployments pay. In that market the advantage goes to buyers who demand operating data, not to those who move first.
This Fortnight
Bedrock Robotics raised $600 million at a $4.1 billion valuation and named eight more contractors for its driverless excavators. It disclosed no machine counts, intervention rates, utilization or cost per cubic yard. Commentators noted that its headline 1.2 million cubic-yard figure is the size of a project, not verified autonomous output. If you run heavy civil work, the customer list makes a pilot worth discussing, but tie the terms to cost per yard moved.
Boston Dynamics launched a second-generation Stretch trailer-unloading robot designed to cut field downtime. It quotes up to 1,000 cases an hour, a vendor figure. DHL's chief executive says use is moving beyond pilots but that polybags remain hard. Treat the rate as a ceiling and test it on your own freight mix, especially soft and bagged parcels.
In an Exotec survey of 250 US and Canadian supply chain leaders, 65% said total cost of ownership came in above expectations. One in four said they wished they had made a different automation choice. The survey comes from a vendor, but the message is clear: budget for integration, exception handling and maintenance before the hardware quote drives the decision.
On October 26 a Starship delivery robot crossed a Tallinn junction on a red light, and a bus braked hard, injuring up to nine standing passengers. Starship blamed darkness and an illuminated advertisement behind the signals and said the fixes would be mainly software. Outdoor robots carry incident risk that also falls on the merchant and platform names on the box, so settle who carries liability before you join a pilot.
In a controlled study, 19 surgeons accepted 60.6% of deliberately misplaced AI guidance overlays and grew more confident after using the AI. The overlays marked incision lines on still images from robotic gallbladder surgery, and some were shifted toward unsafe anatomy. Any operation that relies on human-in-the-loop review (a person checks each AI output before it is acted on) as its safety net should test whether reviewers actually catch errors.
Coming Up
A pedestrian-injury lawsuit against Starship over a 2023 incident at Arizona State University goes to a jury trial in February 2027. Starship argues the pedestrian was negligent. The verdict will be an early marker of where liability lands when sidewalk robots hurt people. If you contract delivery-robot services, review indemnity and insurance terms before then.
FedEx, which moved its robotic trailer loaders into production at its Hagerstown hub in July, plans to extend them to more than 20 US hubs. Its CFO's 2029 targets still assume no material gains from that automation. Expect a wave of vendor case studies from FedEx, DHL and others over the next year. Judge them on completion rates, uptime and cost per unit moved, not robot counts.
Autodesk is moving design and coordination work into its Forma platform and has begun charging for machine-driven compute and data access. Autodesk's chief executive has suggested Forma could match Revit's architectural capability in about two years, and Autodesk has not made clear how it will license existing Revit automation. Construction and engineering firms should list their Revit automations and price the new charges before their next renewal.
What's Hard About This
Accuracy measured in the lab keeps falling apart in the field. An independent study on nine German dairy farms found that only 39.2% of cows flagged by a commercial mastitis sensor actually had clinical mastitis. Vendor benchmarks will not tell you this, so run pilots on your own sites against your own baseline.
Keeping robots busy decides whether they pay. Buying them is the easy part. Serve Robotics grew its delivery fleet to 917 daily active robots in the quarter to September, yet revenue fell, after it said most orders were blocked by restaurant system integration. Most of the return depends on the integration work around the robot, not the robot itself.
Vendors disclose volume but rarely the numbers a buyer needs. GXO said its humanoid robot handled more than 100,000 boxes at a Georgia warehouse but gave no completion rate, intervention frequency or cost per box. Make disclosure of intervention rates and unit costs a condition of any contract.
Practices in this Domain (27)
Read the full technical briefing (1,548 words) →
Where AI Stands in Physical AI & Robotics
Physical AI is where the gap between what a machine can do and what an organisation can deploy is measured most exactly, because failure here has mass and liability. At the mature end the scale is now large. Locus Robotics says it runs more than 22,000 robots across 360 fulfilment centres. Amazon passed its millionth robot in July 2025. Intuitive placed 468 da Vinci systems in the second quarter of 2026. Deere says See & Spray treated more than 5 million acres in 2025, and Brain Corp reports more than 50,000 autonomous cleaning robots. These successes share three conditions: structured floors, narrow repeatable tasks, and a human in control or close by. Where all three hold, the robots work and the returns are documented.
Beyond that core the picture is consistent, and it is not flattering. Most of the domain has working systems, funded vendors and named customers, yet the deployments do not generalise. Only about 6% of North American warehouses are highly automated, a share that has not moved meaningfully since 2024. US precision-agriculture adoption stands at 27% after 30 years on the market. Anthropic's robot exposure index finds robots cost-competitive with human labour for just 0.3% of job tasks. On past price trends, it estimates that share would take 40 years to reach 10%. The model is seldom the bottleneck. More often it is integration with legacy systems, exceptions the robot cannot handle, lifts and doors, alarm floods, payers who will not pay a premium, and liability nobody has settled. Hospital delivery robots stall at elevators more often than they lose their way. Autoliv's diagnostic bot was highly accurate and reached only 20–30% adoption.
This domain stands out because its evidence base is getting more honest about the field gap. Researchers increasingly publish the drop as well as the benchmark. Fully autonomous endovascular navigation fell from 81% success in simulation to 23% in a phantom. A thermal mastitis classifier fell from 82.2% to 60.0% on an unseen farm. Momentum is real in simulation-trained manipulation, construction autonomy and city delivery fleets, where capital and deployments are both growing. Progress has stalled where the claim is autonomy or benefit, not assistance. No autonomous surgical procedure has yet been performed in a human. Companion robots still lack convincing independent evidence that they help psychologically. Whole-home orchestration still breaks routines. Again and again, shared control, narrow scope and hybrid human-machine designs beat full autonomy. The robots that succeed are the ones least ambitious about replacing judgement.
What's New, 2026-10-26 to 2026-11-09
The fortnight was dominated by capital arriving faster than operating data. Bedrock Robotics raised $600M at a $4.1bn valuation and named eight further contractors across eight US states for its autonomous excavators. It disclosed no machine counts, intervention rates, utilisation or cost per cubic yard. Commentators pointed out that its headline 1.2M cubic-yard Zachry figure is the project's scope, not verified autonomous output. They also noted that roughly $1.6B of 2026 construction-robotics funding, more than four times 2025's total, has gone mainly to three firms. Boston Dynamics launched Stretch Gen 2, quoting up to 1,000 cases an hour and more than 20 million boxes unloaded since 2023, all vendor figures. DHL's chief executive said use is moving beyond pilots but that polybags remain hard. Warehouse adoption data cut against the momentum. Link Logistics put highly automated North American sites at about 6%, flat since 2024. An Exotec survey of 250 supply-chain leaders found one in four regretted their automation choice, and 65% said total cost of ownership came in above expectations. In delivery, a Starship robot crossed a Tallinn junction on red on 26 October. A bus braked hard and up to nine standing passengers were injured; Starship blamed darkness and an illuminated advertisement behind the signals. Serve launched on Gold Coast pavements with DoorDash, serving a single restaurant. In the quarter to September its fleet grew to 917 daily active robots while revenue fell to $2.835M against a $59.23M loss.
The research news centred on autonomy and oversight. In a Surgical Endoscopy study, 19 surgeons accepted 60.6% of deliberately misplaced AI overlay suggestions and grew more confident after AI exposure, which undercuts the assumption that surgeon review catches algorithmic error. Simulation-trained robots showed both extremes. Apple researchers reported a humanoid trained only in simulated homes reaching 91% success across 360 real trials, and a Unitree G1 trained only on simulated data beat real-data training by 20 points. On a paper-cup pouring task, by contrast, the ManipTrans baseline fell from 71.7% success in simulation to 0% on hardware. An independent field study on nine German dairy farms found only 39.2% of cows flagged by smaXtec's mastitis alerts actually had clinical mastitis. Restaurant automation also looked weaker. Franchise Times reported Miso's Flippy installed base down to around 14 robots by end-2025, and recorded that Sweetgreen sold its Spyce robotics unit in November 2025 even though its Infinite Kitchen had posted a 31.3% restaurant margin.
Key Tensions
Lab results that do not survive the field. Accuracy measured in the lab or on benchmarks keeps collapsing in deployment, and much of the published evidence is weaker than it looks. A review of 60 strawberry-detection studies found 35% leaked augmented data across train and test splits, inflating test scores by up to 19.28 points, and only 3.3% shared code or weights. Shared-control endovascular navigation reached 92–97% success where full autonomy managed 23% on the same phantom.
Capital outrunning disclosure. Valuations and fleet announcements are growing faster than the operating data that would let a buyer judge them. GXO disclosed that its Agility Digit humanoid had handled more than 100,000 boxes, without a completion rate, intervention frequency or cost per box. JD Logistics claimed 99.9% parcel-tracking accuracy with no stated methodology, and Hai Robotics' IPO filing shows RMB 3.52bn of cumulative losses since 2023. Independent reviewers now routinely advise buyers to run controlled pilots before trusting vendor-wide claims.
Adoption concentrated among the largest operators. Many of the technologies work, but the economics favour operators big enough to spread fixed costs and absorb long integrations. Eurostat data show the 18% of EU farms using precision tools manage 44% of farmland, while variable-rate techniques reach only 5% of farms. Deere has pushed its recurring-revenue goal beyond 2030, citing slower adoption. Robotic surgery earns no payment premium over keyhole surgery, so hospitals adopt it selectively.
Human oversight assumed rather than engineered. Most deployments rely on a person as the final safeguard, and the evidence that this works is thin. Surgeons accepted 60.6% of deliberately wrong AI overlays, and a survey of 72 executives found 96% keep humans reviewing compliance-sensitive AI output, with none running it fully autonomously. Ford rehired 350 veteran engineers after 900 AI cameras missed defects, then moved to hybrid governance.
Utilisation, not robot count, sets the economics. Buying robots is the easy part; keeping them busy and integrated is what decides whether they pay. Serve's fleet grew while its revenue fell, after it said two-thirds of orders were blocked by restaurant back-of-house integration. Hospital delivery routes stall at lifts and doors rather than on navigation. Two Australian home-energy pilots found installs taking nearly a day and households juggling more than six apps.
Top 10 Evidence Items
- Warehouse Automation in 2026: What’s Real vs. Hype | Link Logistics (industry-report) — Anchors the briefing's core claim that mature-domain success hasn't spread: 6% automation share flat since 2024 despite capital inflows. https://www.linklogistics.com/news-insights/original-research/innovation-series/
- AI Construction: Can Driverless Excavators Scale? (opinion) — Exemplifies capital outrunning disclosure — Bedrock's headline cubic-yard figure is project scope, not verified autonomous output. https://www.youtube.com/watch?v=L9EBjoLtQN0
- Clinician-AI Collaboration in Endovascular Neurointerventions: In-Vitro User Study of Copilot Assistance Modes for Robotic Navigation Across the Autonomy Spectrum (research-paper) — The sharpest lab-to-field collapse in the cycle, directly illustrating why full autonomy loses to shared control. https://arxiv.org/html/2611.03446v1
- The surgeon in the loop: automation bias and human–AI interaction in robotic surgery (research-paper) — Undercuts the assumed safeguard of human oversight that the whole domain leans on for compliance and safety. https://link.springer.com/article/10.1007/s00464-026-13440-x
- Serve Robotics Q3 2026: 917 Robots, $2.8M Revenue (adoption-metric) — Shows utilisation and integration, not fleet size, as the binding constraint — robots grew while revenue fell. https://physicalailab.io/en/serve-robotics-q3-2026-results-daily-active-robots-revenue/
- Delivery Robot Runs Red Light in Tallinn, Injuring Bus Passengers (news-coverage) — A real-world liability event that turns the oversight and safety-case debate from theoretical to concrete. https://shiftdelete.net/en/delivery-robot-bus-crash
- Performance of a Sensor-Based Health Monitoring System for the Detection of Clinical Mastitis in Dairy Cows (research-paper) — Field validation outside the vendor's own data shows a flagship livestock sensing product failing most of the time it alerts. https://www.mdpi.com/2076-2615/16/21/3369
- Avoid Robot Regret: What Separates Automation That Delivers from Automation That Disappoints (industry-report) — Buyer-side survey evidence that the widely touted ROI case doesn't hold up once deployed. https://www.exotec.com/insights/avoid-robot-regret/
- Deep learning for real-time strawberry detection, ripeness classification, and picking point localization: a review of architectures, field studies, and open challenges (research-paper) — Exposes how much published robotics evidence is inflated by data leakage, reinforcing the domain's credibility problem. https://link.springer.com/article/10.1007/s10462-026-11747-8
- Boston Dynamics designs Stretch Gen 2 to maximize uptime (product-ga) — Illustrates the pattern of headline vendor figures (uptime, boxes unloaded) substituting for independently verified performance data. https://www.automatedwarehouseonline.com/boston-dynamics-designs-stretch-gen-2-maximize-uptime/