Erik Caldwell
Connect on LinkedIn
Menu

The Topic Catalog · 21

Physical AI & Autonomous Systems

Last updated · 35 min read

Based on research by Erik Caldwell, "When AI Enters the Physical World: Robotics, Autonomous Systems, and the Future of Work" (October 2026). Evidence in this domain is current to 16 October 2026. Figures reported only by the company that sells the product are labeled as company-reported; maturity judgments and the scoring framework in 21.5 are analyst judgments anchored to evidence, not measurements.

21.1 Physical AI Defined: What Is Learned, What Is Engineered, and the Five-Gate Deployment Gap

Priority: Must Understand

Executive Definition: Physical AI is machines and systems that perceive their physical surroundings, make decisions with some degree of independence, and take physical action, where at least part of the perception or decision-making is learned from data rather than fully specified by engineers. It spans warehouse robots, autonomous vehicles, drones, factory arms, inspection robots, and humanoids. Much of the most valuable autonomy is not learned at all: autonomous mining haul trucks have run for nearly two decades on largely classical navigation and fleet management, and ocean floats operate for months on deterministic control.

Why It Matters: Software AI can usually skip the hard gates because its errors are cheap and reversible. Physical AI cannot, because its errors injure people and break things. A physical AI system has to clear five gates: capability, reliability, assurance, economics, and acceptance. Most programs that failed between 2022 and 2026 cleared the first gate and stalled at a later one: Cruise lost its permits after a single mishandled incident, Kroger closed three of eight Ocado automated fulfillment centers and took a $2.6 billion charge because the economics did not hold, and Amazon shelved its Blue Jay warehouse robot within six months over cost and integration problems. An organization that evaluates physical AI on demonstrations will consistently overestimate how quickly it will pay.

What I Need to Understand:

  • "Autonomous" and "AI-driven" are different claims. Many scaled systems use classical control or task-specific computer vision, and calling them "AI" confuses the question, while dismissing them misses the lesson that simple, verifiable autonomy often beats sophisticated learning in hard environments.
  • The vocabulary matters because vendors blur it: industrial arms (programmed, mostly classical control), collaborative robots or cobots (arms built to work near people under force and speed limits), autonomous mobile robots or AMRs (wheeled robots navigating facilities), mobile manipulators (a mobile base plus arms), humanoids, autonomous vehicles, uncrewed aircraft, maritime autonomy, and teleoperated or supervised robots where a human holds authority. Robot foundation models, vision-language-action models, and world models are the learned-intelligence layer that can sit on top of several of these bodies.
  • The five gates are different tests. Reliability asks how often the system fails or needs a human, measured in intervention rates and mean time between failures, not best-case success rates. Assurance asks whether safety can be demonstrated to a regulator, insurer, union, or city council with evidence a third party can check. Economics asks what a productive hour costs once integration, maintenance, supervision, downtime, and depreciation are included, compared with both human labor and conventional automation. Acceptance asks whether workers, the public, procurement officers, and elected officials will allow it.
  • Treat a deployment claim by its evidence level: laboratory research, controlled demonstration, field pilot, limited commercial deployment, scaled commercial deployment. A signed partnership, purchase agreement, memorandum of understanding, or announced plan is not a deployment.
  • Scaled deployment lives where the environment is engineered and the task is narrow. Robotaxis are the exception that proves the rule: they reached roughly half a million paid rides a week only after more than fifteen years and only inside mapped, geofenced cities.
  • Capital is a signal of expectation, not of performance. Agility Robotics' public listing implied a valuation of roughly 1,400 times its 2025 revenue, and Unitree's market value peaked near $66 billion on its listing day against roughly $250 million in annual revenue.

Questions I Should Be Able to Ask My Team:

  1. For this proposed system, which part is learned and which part is conventional engineering, and does the vendor's marketing describe the same split?
  2. Which of the five gates has this vendor actually cleared in an environment like ours, and what is the evidence for each one?
  3. Are we evaluating this on a demonstration, or on operating data such as intervention rates, uptime, and cost per productive hour from comparable sites?

Technologies / Standards / Companies to Know: Industrial robots, cobots (ISO/TS 15066), autonomous mobile robots (AMRs), mobile manipulators, vision-language-action (VLA) models, geofenced operating domains, intervention rate, robots-as-a-service.

Recommended Learning:

Time Investment: 1 hour

21.2 What Works Today: Narrow Autonomy at Scale

Priority: Must Understand

Executive Definition: The physical AI that is paying its way in 2026 is narrow, supervised autonomy in engineered or bounded environments: autonomous mining haulage, structured warehouses, geofenced robotaxis, drone delivery, vision-guided precision agriculture, industrial inspection, and ocean observation. These systems succeed because the environment is controlled, the task is narrow, the economics are measurable, and humans still supervise the exceptions.

Why It Matters: The popular picture of physical AI is a humanoid video. The operating record is different. As of October 2026, Waymo delivers roughly 500,000 paid driverless rides a week across 15 US cities, Komatsu has commissioned its 1,000th ultra-class autonomous haul truck, Amazon operates more than one million warehouse robots, Zipline has passed two million commercial drone deliveries, and John Deere's camera-guided sprayers covered more than five million acres in the 2025 season. Almost none of these systems relies on a general-purpose foundation model for its core function. The scaled present of physical AI is mostly narrow AI and good engineering, and that is the most important correction to make before setting an investment thesis.

What I Need to Understand:

  • The programs that reached scaled commercial deployment share five traits: a structured or engineered environment, a narrow and repetitive high-frequency task with clear unit economics (herbicide per acre, picks per hour, tonnes hauled, rides per week), a human-in-the-loop or assistive design, an anchor customer who absorbs integration risk (Walmart for Symbotic, FedEx for Dexterity, DHL for Locus and Stretch), and a business model that tolerates slow payback. Robots-as-a-service is now the dominant model in the US professional service robot market, according to the International Federation of Robotics (IFR).
  • Waymo has the strongest published safety record in the field: over 56.7 million miles it reported 85% fewer serious-injury crashes than a human benchmark, and a Swiss Re study found 88% fewer property-damage claims. Both analyses were authored or co-authored with Waymo, so read them as the best available evidence rather than independent proof.
  • Where AI actually sits in scaled systems is often a layer on top of conventional operations. Amazon's DeepFleet model coordinates fleet traffic and, by Amazon's own account, cut robot travel time by 10%; that figure is company-reported and not independently verified.
  • Supervision does not vanish, it moves. Remote assistance, in-cab observers, operations centers, and exception handling are the new labor of autonomy, and organizations should design for them rather than assume them away.
  • One incident can freeze a market. Cruise lost its permits in 2023 after a mishandled incident, and a mass system failure in Baidu's Wuhan robotaxi fleet in March 2026 stalled service citywide and led China to pause new robotaxi permits nationally. Fleet-level software risk is systemic, and regulators respond at the level of the whole market.
  • The same programs also show what does not scale. Ocado's customer Kroger closed three of eight automated fulfillment centers when demand shifted, Monarch Tractor collapsed after dealers alleged its tractors could not operate autonomously, and Amazon Prime Air has completed only about 16,000 deliveries against an original goal of 500 million a year by 2030.

Questions I Should Be Able to Ask My Team:

  1. Which of our candidate tasks look like the scaled cases (structured, narrow, high frequency, measurable), and which only resemble the demo?
  2. If this fleet stopped tomorrow because of a vendor-wide software failure, what would we do operationally, and how long could we operate without it?
  3. Who is the anchor customer or comparable site for this vendor, and have we spoken to them rather than relying on the vendor's reference list?

Technologies / Standards / Companies to Know: Waymo, Baidu Apollo Go, Komatsu, Caterpillar, Amazon Robotics, Symbotic, Locus Robotics, Zipline, Wing, John Deere See & Spray, Carbon Robotics, Intuitive (da Vinci teleoperated surgery).

Recommended Learning:

Time Investment: 1 hour

21.3 Robot Foundation Models, VLAs & the Data Problem

Priority: Should Understand

Executive Definition: Robot foundation models are large models pretrained on diverse data, typically web data plus data from many different robots, and then adapted to specific tasks and bodies. Vision-language-action (VLA) models are the main form: they take camera images and language instructions and output motor actions. The shift they represent is from one robot, one task, trained from a few hundred demonstrations, to pretrain once and adapt many.

Why It Matters: This is the most significant technical shift in robotics in decades, and it is real. Google DeepMind's RT-2 (2023) showed web knowledge transferring into robot control, the Open X-Embodiment collaboration pooled data from 22 robot types and found that training on other robots' data improved each robot, and Physical Intelligence's π0.5 cleaned kitchens and bedrooms in homes it had never seen, with web data accounting for the difference between 94% and 74% success on unfamiliar objects. But the field has no shared benchmark, reliability is far from industrial standards, touch and force are largely missing from training data, and the robot data gap relative to language models is enormous. Robotics is having its GPT-2 or GPT-3 moment for research. It is not yet having its ChatGPT moment for deployment.

What I Need to Understand:

  • Improvements in general multimodal models transfer strongly to semantics, perception, planning, instruction following, and checking whether a task succeeded (Google DeepMind's Gemini Robotics-ER reads industrial gauges with 86% to 93% success, up from 23% in the prior version). They transfer partially to generalization across environments. They do not yet transfer to contact-rich control, force regulation, long-horizon reliability, safety guarantees, or hardware durability, which still need robot-specific data, mechanisms, and engineering.
  • Data is the binding constraint. Ken Goldberg estimated in Science Robotics that the text and images behind vision-language models amount to roughly 100,000 years of human reading and viewing, against about one year of robot experience in the largest robot dataset, a rough extrapolation but the right order of magnitude. The field disagrees about how to close the gap: real deployment flywheels (Sergey Levine), simulation first (NVIDIA), or human video (Skild AI). One careful study found that generalization follows the number of distinct environments and objects, not the number of demonstrations.
  • Evaluation is the field's weakest link. Most results are success rates on the authors' own hardware, tasks, and scenes, often over tens of trials, and one researcher raised a task's success from 5% to 95% by rewording the instruction. Toyota Research Institute's blind, randomized evaluation of Large Behavior Models and the crowd-sourced RoboArena comparisons are the credible exceptions.
  • Read every reported success rate with five questions: whose hardware, seen or unseen environments, how many trials, whether humans reset or coached the robot, and how throughput compares with a person.
  • Intelligence is decoupling from bodies. Cross-embodiment models and autonomy stacks are becoming portable across hardware, which means value and lock-in will concentrate in software and operational data.
  • Operational data from real deployments is becoming a strategic asset. Contracts must define who owns it, how long it is retained, and whether it is portable if the vendor exits.
  • Sim-to-real transfer is mature for legged locomotion and unsolved for dexterous manipulation, and world models are promising as generators of interactive environments but not yet demonstrated as physics-faithful training environments for manipulation.

Questions I Should Be Able to Ask My Team:

  1. When a vendor reports a success rate, was it measured on their hardware or ours, in seen or unseen environments, over how many trials, and were humans resetting or coaching the robot?
  2. Who owns the operational data our deployed machines generate, and could we take it with us if the vendor failed?
  3. Does the vendor's model improve from our deployment, and what do we receive when it does?

Technologies / Standards / Companies to Know: RT-2, OpenVLA, Open X-Embodiment, Physical Intelligence (π0 family), Google DeepMind Gemini Robotics, NVIDIA GR00T and Cosmos, Figure Helix, Skild AI, Toyota Research Institute Large Behavior Models, RoboArena.

Recommended Learning:

Time Investment: 2-3 hours

21.4 Humanoids & General-Purpose Robots

Priority: Monitor

Executive Definition: Humanoid robots are two-armed, often bipedal machines built to operate in spaces designed for people, with learned whole-body control and manipulation at the center of the design. The argument for them is that the world is built for human bodies, so a human-shaped robot can in principle do any job.

Why It Matters: Humanoids absorb a large share of the attention and capital in physical AI and a very small share of the operating record. The IFR counted about 7,000 full-size humanoids sold worldwide in 2025, against more than 600,000 industrial robot installations and about 250,000 professional service robots, and describes a "weak business case, especially in industrial settings." Most humanoid buyers are research labs, universities, and data-collection programs. The best-documented productive humanoid deployment, Agility Robotics' Digit, reported 65,000 operating hours across nine customer facilities and $1.8 million in 2025 revenue. On current evidence humanoids will not dominate physical AI before the mid-2030s, and wheeled, legged, fixed, aerial, and marine systems will carry most of the economic value.

What I Need to Understand:

  • General-purpose robots are not outperforming specialized machines on any verified metric. The best-documented productive humanoid tasks, moving totes, loading sheet metal, and sequencing parts, are tasks that conveyors, AMRs, and fixed automation already perform, and no buyer has published a side-by-side comparison of cost, throughput, or uptime. Within one company the contrast is plain: Boston Dynamics' wheeled, single-purpose Stretch has been in commercial use at DHL since 2023, while its Atlas humanoid is not scheduled for productive use in Hyundai plants until 2028.
  • Bodies are not on a software cost curve. Figure AI's own account of eleven months at BMW identified the forearm as the top hardware failure point. Unit prices are falling (Unitree's average humanoid selling price fell from about $85,000 in 2023 to about $25,000 in 2025, and Morgan Stanley estimates about $150,000 in 2028 falling to about $50,000 by 2050), but cheaper platforms do not by themselves prove labor productivity. Batteries, actuators, hands, thermal management, and field serviceability remain engineering problems, not model problems.
  • Safety near people is unresolved. A falling full-size biped carries significant kinetic energy, and the first ISO standard aimed at dynamically stable industrial robots, ISO/CD 25785-1, is still a committee draft.
  • Founder timelines have been early by years. Tesla's public Optimus volume targets for 2025 were missed, and Rodney Brooks predicts profitable dexterous humanoid deployment is more than ten years away. A useful rule is to stretch founder and CEO deployment timelines by two to three times, and to give more weight to forecasters who publish dated, self-graded predictions.
  • Where generality may pay off first is software, not bodies: cross-embodiment models may make specialized machines more adaptable well before humanoids become economical.
  • What would change this assessment: customers (not vendors) publishing cost-per-productive-hour comparisons in which humanoids beat both labor and conventional automation, a practical method for collecting touch and force data at scale, and publication of ISO 25785-1 with a certification pathway for learned control near people.

Questions I Should Be Able to Ask My Team:

  1. Has any customer, not the vendor, published cost per productive hour, throughput, and uptime for this task against conventional automation?
  2. Could a wheeled, fixed, or conventional system do this task today, and what does the humanoid form factor actually add?
  3. If this vendor consolidates or fails, what are we left with: stranded hardware, a redesigned facility, or an exclusive contract?

Technologies / Standards / Companies to Know: Agility Robotics (Digit), Figure AI, Tesla Optimus, Boston Dynamics (Atlas, Stretch), Unitree, AgiBot, UBTech, ISO/CD 25785-1.

Recommended Learning:

Time Investment: 1 hour

21.5 Buying Physical AI: Evidence Standards & the Premature-Deployment Test

Priority: Must Understand

Executive Definition: Evidence-based procurement of physical AI means requiring operating evidence, not demonstrations, before committing: intervention rates, uptime, mean time between failures, and cost per productive hour from environments like yours, plus contract terms that protect data rights, update paths, and exit options. The one question to ask every vendor is: what is your cost per productive hour, at what intervention rate, in an environment like ours?

Why It Matters: Physical AI projects fail the same way enterprise IT and the Internet of Things did, and for organizational reasons more than technical ones. Cisco's 2017 survey found 60% of IoT initiatives stalled at proof of concept, with IT and business collaboration, executive sponsorship, and expertise as the top success factors. The Government Accountability Office's review of Healthcare.gov found task orders issued before requirements were known and readiness verification skipped. The integration burden for physical AI is heavier, because facilities, safety cases, maintenance, unions, insurers, and regulators all become stakeholders. Vendor performance figures are mostly company-reported, so the buyer has to ask for the data that is rarely published.

What I Need to Understand:

  • Four principles organize the work: buy evidence, not demos; prefer no-regret investments; govern before deploying; and keep options open by favoring modular architectures that separate the intelligence layer from the hardware.
  • The metrics that matter are rarely public today: intervention rate (human assists per hour or mile), mean time between failures and maintenance cost, task success on unseen sites, time and data needed to learn a new task, supervision ratio (machines per remote operator), repeat purchases by named customers (second and third orders, not memoranda of understanding), production volume against capacity claims, and safety incidents per unit of exposure.
  • A simple scoring framework helps structure the decision: environmental structure (15 points), task narrowness and repeatability (15), cost of error (15), evidence of reliability (15), unit economics (15), organizational readiness (10), acceptance and governance (10), and vendor durability and openness (5). A total of 75 or more justifies a time-boxed pilot, 55 to 74 means fix the weak criteria first, and below 55 means monitor. Any single score of 1 on cost of error, reliability, or governance should stop the project regardless of the total.
  • Treat a deployment as premature if the vendor cannot provide intervention rates or uptime from a comparable environment, performance was measured only in the vendor's own facility, the business case assumes headcount reductions the vendor's own customers have not achieved, the task is unstructured or high-consequence with no minimal-risk fallback, the system will collect data about the public with no approved use policy, the hardware has supply-chain exposure or no credible update path, nobody inside the organization owns safety, maintenance, and model updates, workers were not consulted, the vendor's revenue is small against its cash burn with no exit or data-portability clause, or success is defined as "deploying robots" rather than an improved outcome.
  • Total cost of ownership, not unit price, decides value. Robots-as-a-service shifts spending from capital to operating budgets and blends hardware and software contracting, which most procurement rules were not written for.
  • Every pilot needs success criteria, a time limit, and an exit plan.

Questions I Should Be Able to Ask My Team:

  1. What evidence would we require before signing for a robot or autonomous system, and do our procurement rules actually require it?
  2. Does the contract give us ownership of, or rights to, the operational data, update and patch rights, and portability if the vendor exits?
  3. What are this pilot's success criteria, time limit, and exit plan, and who decides whether it scales?

Technologies / Standards / Companies to Know: Robots-as-a-service (RaaS), total cost of ownership, minimal-risk condition, operational design domain, software bill of materials (SBOM), outcome-based payment.

Recommended Learning:

Time Investment: 2-3 hours

21.6 Safety, Standards & Liability for Learned Control

Priority: Should Understand

Executive Definition: Safety assurance for physical AI is the discipline of demonstrating, with evidence a third party can check, that an autonomous machine is safe to operate near people, including how it behaves when it loses communications, meets something new, or is updated. The existing functional safety standards were written for deterministic machines that can be caged, fenced, or monitored by speed and separation.

Why It Matters: No certification framework for end-to-end learned control was found in the research behind this domain. The ISO 10218 industrial robot standards were republished in February 2025, and ISO 3691-4 covers driverless industrial trucks, but the first ISO standard for dynamically stable industrial robots, including humanoids and legged machines, closed its committee-draft comment period in July 2026 and is not published. Standards are voluntary unless referenced in law, and liability is moving ahead of them: the EU's revised Product Liability Directive explicitly covers software and AI and applies from 9 December 2026. Until certification exists, near-term deployments will rely on engineered containment, speed limits, zoning, and human oversight around the learned components.

What I Need to Understand:

  • The standards landscape: ISO 10218-1 and 10218-2 (industrial robots and cells), ISO/TS 15066 (collaborative robots, under revision), ISO 13482 (personal care robots, near a new edition), ISO 3691-4:2023 (driverless industrial trucks), and ISO/CD 25785-1 (dynamically stable industrial mobile robots, draft). None certifies learned end-to-end control.
  • Liability differs by jurisdiction. The EU Product Liability Directive can attach liability to defects that emerge through updates or machine learning. The UK's Automated Vehicles Act places liability on manufacturers when a vehicle is in self-driving mode, but is enacted and not yet in force, with full implementation expected in late 2027. In the US, liability remains largely state tort law, with federal preemption proposals pending.
  • Defense assurance practice is the most fully specified public framework for autonomous systems, and some of it transfers. US Department of Defense Directive 3000.09 requires operational testing in realistic conditions against adaptive adversaries, behavior that terminates or seeks operator input when limits are reached, re-verification when software changes so safety features are not degraded, new senior review when algorithms, missions, or environments change substantially, and human-machine interfaces that make clear what the operator does and what the system does.
  • What transfers to civilian use: realistic operational testing including red-teaming, required behavior on communications loss or exit from the operating domain (a minimal-risk condition), re-certification triggers for software and over-the-air model updates, and trust calibration as a measured objective. What does not: attritable risk tolerance, defense cost structures, secrecy, military accountability chains, and urgent-need waivers of review.
  • Connected machines are cybersecurity assets. The EU Cyber Resilience Act brings connected robots under product security duties, with vulnerability reporting since September 2026 and full obligations from December 2027. The FCC's 2026 robot action allows mitigating software updates to already-authorized robots only until 1 January 2029, which creates a concrete update-path risk for existing fleets.
  • A model update is a safety event. Safety, cybersecurity, and model-update control need one accountable owner, because operational technology, IT, and AI governance converge in the same machine.

Questions I Should Be Able to Ask My Team:

  1. For every autonomous system we run, who is accountable for its safety, its cybersecurity, and its software and model updates?
  2. What does each system do when it loses communications or leaves its operating domain, and has that behavior been tested rather than assumed?
  3. When the vendor ships an over-the-air update, what triggers re-verification before it reaches our site, and can we delay or reject it?

Technologies / Standards / Companies to Know: ISO 10218, ISO/TS 15066, ISO 13482, ISO 3691-4, ISO/CD 25785-1, DoD Directive 3000.09, EU Product Liability Directive, EU Machinery Regulation (applies 20 January 2027), EU Cyber Resilience Act, UK Automated Vehicles Act 2024.

Recommended Learning:

Time Investment: 2-3 hours

21.7 Supply-Chain Law, Geopolitics & Regulation

Priority: Should Understand

Executive Definition: Which robots and drones an organization may buy and operate, and which jurisdictions' rules apply, is increasingly set by trade, security, and product law rather than by technology. Governments relate to physical AI in three roles that are easy to conflate: as users that buy and operate autonomous systems, as regulators that set safety, liability, privacy, and labor rules, and as industrial strategists that fund research, subsidize adoption, and protect supply chains.

Why It Matters: Geopolitics now shapes physical AI options as much as technology does. The United States has no national robotics strategy and acts mainly through security and trade law: the FCC added foreign-produced drones to its Covered List in December 2025 and foreign-produced ground robots in July 2026, and Section 301 tariffs of 25% already apply to Chinese robotics. China installs 59% of the world's industrial robots (354,000 in 2025), has made embodied AI a priority in its 15th Five-Year Plan, and leads humanoid shipments. The EU has the most complete legal framework and is re-timing it. A buyer that ignores this can find that a product it standardized on cannot be replaced or expanded. Status below is as of 16 October 2026.

What I Need to Understand:

  • The FCC's July 2026 action blocks new equipment authorizations for foreign-produced ground robots over 4.4 pounds with sensors, connectivity, and control software, including humanoids, quadrupeds, warehouse robots, and sidewalk and inspection robots. "Foreign-produced" is defined by a Buy American test of 65% domestic component cost, rising to 75% in 2029, and new models need a Department of War conditional approval. Existing fleets can keep operating, but replacement and expansion will require compliant hardware, likely at higher unit cost.
  • Enabling rules have lagged restrictive ones. The FAA's rule for drone flights beyond visual line of sight (Part 108) missed its deadlines and was not final at the cutoff, OSHA has no robotics-specific standard, and NHTSA has proposed rewrites of federal motor vehicle safety standards for autonomous vehicles.
  • The EU delayed high-risk AI Act obligations through the Digital Omnibus, enacted July 2026: stand-alone high-risk systems to 2 December 2027 and product-embedded systems such as machinery to 2 August 2028. The Machinery Regulation applies from 20 January 2027, before the dedicated ISO safety standard for humanoid and legged robots is published, which creates a sequencing problem for robot makers.
  • China is the clearest industrial strategist, combining a national plan, very large public capital, local subsidies, and demonstration zones, and it has begun to manage overheating: its planning agency warned of a humanoid bubble in November 2025 and paused new robotaxi permits nationally after the Wuhan outage.
  • South Korea, Japan, and Singapore act mainly as users driven by demographics, commissioning autonomous buses and funding care robots, with modest results so far: small ridership, weak fare recovery, safety staff still aboard, and care robots that complement rather than replace staff.
  • Procurement implications: check Covered List status, domestic content, and export controls in every solicitation, require update and patch rights, and plan for vendor consolidation and higher replacement cost.

Questions I Should Be Able to Ask My Team:

  1. Is any hardware in our fleets or purchase plans foreign-produced under the FCC's Covered List actions, and what is our replacement and update path if it is?
  2. How exposed are we to tariffs, export controls, or trade restrictions on actuators, sensors, batteries, and compute in the machines we buy?
  3. Do our solicitations check Covered List status and domestic content, and do our contracts survive a vendor exit or acquisition?

Technologies / Standards / Companies to Know: FCC Covered List, Section 301 tariffs, Section 232 robotics investigation, FAA Part 108, EU AI Act and Digital Omnibus, EU Machinery Regulation, China's 15th Five-Year Plan, Korea's K-Humanoid Alliance.

Recommended Learning:

Time Investment: 1 hour

21.8 Workforce, Organization & the Supervision Shift

Priority: Must Understand

Executive Definition: Physical AI reorganizes work before it removes it. The best causal evidence says automation produces productivity gains at adopting firms, declining labor shares, reallocation from production workers toward technical and service roles, and labor adjustment that arrives mainly through reduced hiring of young entrants rather than layoffs. Supervision moves from the vehicle or the machine to the operations center, and routine handling gives way to exception management and maintenance.

Why It Matters: The workforce question for the next decade is less how many jobs disappear than which entry points to skilled work close, and who builds the technician, supervisor, and safety roles that replace them. The evidence base is the literature on traditional industrial robots, mostly in automotive and electronics manufacturing before 2015, and extrapolating it to AI-enabled machines is an assumption, not a finding. No credible job-loss number for physical AI exists. What the evidence supports is a direction (task restructuring, entry-level pressure, technician demand), a mechanism (adjustment through hiring), and a pace bounded by the deployment gap in 21.1.

What I Need to Understand:

  • The strongest findings, with their limits: in US local labor markets from 1990 to 2007, one more robot per 1,000 workers lowered the employment-to-population ratio by 0.2 percentage points and wages by 0.42% (Acemoglu and Restrepo); in Germany, about 2.11 manufacturing jobs were lost per robot, offset by service jobs, with net employment about unchanged and the adjustment coming through fewer young entrants (Dauth and coauthors); across 17 countries, robots raised annual labor-productivity growth by about 0.36 percentage points (Graetz and Michaels). Early generative AI shows the same entry-level pattern: workers aged 22 to 25 in the most AI-exposed occupations saw a 13% relative employment decline (Brynjolfsson, Chandar, and Chen), from one payroll provider and in knowledge work.
  • Think in task bundles. Augmented: maintenance and robotics technicians, reliability engineers, novice workers aided by AI. Restructured: warehouse pickers at robotic stations, security guards supervising patrol robots, truck drivers whose non-driving tasks persist. Displaced, mostly through reduced hiring: machine feeders and packers, forklift operators, entry-level manufacturing hires. New roles: remote operations and assistance staff, fleet supervisors, robot trainers and data collectors, safety and operating-domain validation engineers.
  • Protect the entry-level ladder. If routine physical tasks that once trained new workers are automated, the organization loses the apprenticeship path that produced its future technicians and supervisors. Redesigning entry roles around supervision, exception handling, and maintenance is a workforce strategy, not a social gesture.
  • Safety and productivity are design choices. A 2024 US Senate committee majority staff report on Amazon linked warehouse pace to injury, and investigative reporting found higher serious-injury rates at robotic warehouses in 2016 to 2019. Robots can remove walking and lifting while raising the repetition rate of the human tasks that remain, so whether automation makes work safer depends on how pace is managed.
  • Labor shortages pull adoption, and the choice is managerial. Japan faces a projected shortfall of more than 11 million workers by 2040, while US Bureau of Labor Statistics projections show modest growth, not decline, for heavy truck drivers and for hand laborers and material movers through 2035, with most openings being replacement churn. Automation can address churn and injury without shrinking headcount, or it can quietly reduce hiring.
  • Read forecasts for what they measure. Frey and Osborne's 47% of US jobs at high risk scored whole occupations, while a task-level method yielded about 9%. McKinsey Global Institute's 57% of US work hours is technical potential, explicitly not a forecast of job losses.
  • Organizations will run shared processes that coordinate employees, software agents, and machines, and each hand-off needs a defined competence envelope. In practice: supervisors manage exceptions and fleets, workflow redesign precedes automation (Elon Musk conceded in 2018 that "excessive automation at Tesla was a mistake"), measurement shifts to cost per productive hour, intervention rate, uptime, and safety outcomes, and joint technology review with labor reduces resistance and preserves institutional knowledge.
  • Adoption concentrates in large firms, and non-adopting competitors lose employment, so without shared-access models such as robots-as-a-service, manufacturing extension partnerships, and community college programs, physical AI is likely to widen gaps between large and small firms and between regions.

Questions I Should Be Able to Ask My Team:

  1. If automation removes the entry-level jobs that used to train them, where will our future technicians and supervisors come from?
  2. Have we engaged our workforce and their representatives before, rather than after, deciding what to automate, and what redeployment commitments are we prepared to make?
  3. Is the business case built on measured outcomes or on headcount reductions, and have the vendor's other customers actually achieved those reductions?

Technologies / Standards / Companies to Know: Task-based job analysis, remote operations centers, robots-as-a-service, joint technology review clauses in labor agreements, McKinsey Global Institute and World Economic Forum workforce projections, US Bureau of Labor Statistics Occupational Outlook Handbook.

Recommended Learning:

Time Investment: 2-3 hours

21.9 Governing Physical AI in Public Space: Surveillance, Trust & Local Government

Priority: Should Understand

Executive Definition: Robots, drones, and autonomous vehicles are mobile sensors, so deploying them in public space is judged on purpose and social acceptability as much as on technical safety. Governance, meaning use policies, data rules, oversight, and public commitments, has to precede deployment, and for US cities and counties it is the real gate on adoption.

Why It Matters: For local governments, procurement rules, surveillance-technology ordinances, labor agreements, and public trust will shape adoption more than technology does over the next three years. The history of smart-city technology shows the pattern. San Diego's smart streetlight program, approved in 2016 and pitched as self-funding through energy savings, saw police access its cameras at least 35 times during the 2020 protests, went dark amid a vendor ownership change, and prompted the city's surveillance ordinance after the fact (as reported by Voice of San Diego). Toronto's Sidewalk Labs project was cancelled in 2020 after sustained privacy and data-governance criticism. The same dynamics apply to any organization that deploys machines in spaces shared with the public.

What I Need to Understand:

  • Community-control surveillance ordinances, adopted in San Diego (2022), Oakland (2018), and more than two dozen other jurisdictions, require use policies, impact reports, advisory-board review, council approval, and annual reports before police surveillance technology is deployed. Budget months for that process, publish narrow use policies, and ban weaponization in writing.
  • Public-safety robotics is governed by trust more than technology. Chula Vista's tightly scoped drone program has endured, while the New York and San Francisco police ground-robot episodes became controversies.
  • What US local governments can practically do in the next three years: drones as first responders under a narrow use policy and compliant hardware, AI-assisted wildfire, flood, and infrastructure cameras, pipe, sewer, bridge, and facility inspection, autonomous floor cleaning in city buildings, robotaxi coordination (curb management, emergency-response protocols, data-sharing agreements), and precision grounds maintenance.
  • What to monitor without committing major funds: humanoid robots in public works, autonomous transit buses as a core service, sidewalk delivery robots in the public right of way, general-purpose service robots in libraries, parks, and social services, and fully autonomous refuse collection. Avoid any use-of-force or weaponized robot.
  • The most successful government uses are often the least glamorous: ocean observation, wildfire detection, and infrastructure monitoring, where simple autonomy produces public data and puts no one's job or civil liberties at stake. Public transit autonomy is technically workable but economically marginal at current scale.
  • Local governments face fixed budgets and capital cycles, procurement rules designed for commodities, collective bargaining agreements, public records obligations, tort liability, thin technical capacity, and vendor dependence. Expect unions to seek technology-review clauses like the one in the 2024 to 2030 East and Gulf Coast port contract, and involve them before pilots, not after.
  • Autonomous vehicle interaction is already a local issue: NHTSA has flagged driverless vehicles interfering with first responders, so police and fire departments need incident protocols with operators.

Questions I Should Be Able to Ask My Team:

  1. Do we have an approved use policy, including a written ban on weaponization and a data retention rule, before any robot, drone, or camera purchase?
  2. Who can access the data our machines collect, for how long, and how would we answer a public records request, subpoena, or incident inquiry?
  3. What happens to this service and its data if the vendor changes ownership, changes terms, or exits mid-contract?

Technologies / Standards / Companies to Know: Community-control surveillance ordinances, drone-as-first-responder programs, FAA Part 108 (beyond-visual-line-of-sight), curb management and AV incident protocols, autonomous shuttle licensing.

Recommended Learning:

Time Investment: 1 hour

21.10 Readiness: No-Regret Moves, Timelines & Leading Indicators

Priority: Must Understand

Executive Definition: Physical AI readiness means investing now in capabilities that pay off whether physical AI arrives fast or slow, piloting where structure is high and outcomes are measurable, treating general-purpose robots as options to watch, and tracking a defined set of leading indicators that tell you when to move from monitoring to investing.

Why It Matters: Physical AI advances task by task and environment by environment, fastest where the world can be engineered to suit machines and slowest where machines must adapt to the world. The organizations that benefit most will not be those that bet earliest on the most humanlike robots, but those that learn to measure, govern, integrate, and maintain autonomous systems in their own operations, so that when general-purpose capability matures they are ready to absorb it. Dated timelines have a poor record: Rodney Brooks tallied 20 dated industry predictions about autonomous vehicles through 2025, and 18 had not come true.

What I Need to Understand:

  • Separate five milestones: first technical demonstration, reliable operation in controlled environments, limited commercial deployment, economic viability (positive unit economics without subsidy), and widespread adoption. Waymo operated in three cities in September 2024 and fifteen by September 2026, yet economic viability at scale is unproven for every operator and widespread adoption has not occurred anywhere. The base rate from demonstration to roughly half a million weekly commercial rides has been more than fifteen years.
  • The base case for 2026 to 2040 is steady expansion concentrated where structure and economics are strong: foundation models make specialized machines more adaptable, humanoids find niches in brownfield manufacturing logistics and data collection, robotaxis and driverless trucking grow city by city and lane by lane with periodic regulatory pauses, and household robots remain niche and supervised through the mid-2030s. An accelerated and a constrained scenario bracket it, and the leading indicators below are how you tell which one is unfolding.
  • No-regret investments, useful in every scenario: task-level process mapping and baseline metrics, operational data and sensor architecture with clear ownership, cybersecurity for connected machines with update and patch rights, procurement templates that require reliability and safety evidence, technician, maintenance, and supervisor training pipelines, governance (use policies, safety cases, incident reporting), and pilots in structured, measurable applications.
  • Speculative investments, dependent on specific technologies succeeding: large fleets of general-purpose humanoids, long exclusive contracts with a single autonomy vendor, facility redesigns predicated on humanoid labor, workforce reductions planned on vendor automation-potential figures, bets on a single robot foundation-model provider, and proprietary world-model platforms without field evidence.
  • Three horizons for action. Now (2026 to 2028): map processes at the task level and score tasks on structure, repetition, error cost, and labor scarcity, pilot in logistics, inspection, cleaning, and quality vision, build an operational technology and robotics cybersecurity baseline, establish data ownership in every automation contract, and start a technician pipeline. Prepare (2029 to 2031): redesign workflows around human, software-agent, and robot hand-offs, build a data architecture that captures operational data, negotiate multi-vendor modular autonomy stacks, retrain incumbent staff, and prepare facilities. Monitor (2032 to 2040): general-purpose humanoids in mixed environments, household robots, and autonomous construction at scale.
  • Leading indicators to track: sustained paid operations across several operators (not one leader), customer-published cost per productive hour, intervention-rate disclosure becoming standard, supervision ratios improving, repeat purchases and fleet expansion by named customers, units shipped to productive use versus capacity claims, safety incidents and recalls per unit of exposure, standards milestones such as ISO/CD 25785-1, regulatory approvals and pauses, and labor agreements that address automation.
  • Slower-adoption signals: a fatal or high-profile incident involving a general-purpose robot near the public, a sharp correction in humanoid and robot-AI valuations, repeated fleet-wide software failures like Wuhan's, escalating trade restrictions on components, and labor conflicts that write broad restrictions into contracts or law.

Questions I Should Be Able to Ask My Team:

  1. Which indicators will we track to decide when to move from monitoring to investing, and who reviews them each year?
  2. Have we mapped our operations at the task level, and can we state the baseline cost, quality, safety, and throughput of each candidate task today?
  3. Do we have a written policy for data collected by robots, drones, and autonomous vehicles, covering retention, access, and sharing, and do our contracts give us rights to that data?

Technologies / Standards / Companies to Know: Task-level process mapping, operational technology (OT) cybersecurity, robots-as-a-service, modular autonomy architectures, and the indicators published by Waymo, Baidu, Zipline, the IFR, and NHTSA crash reporting.

Recommended Learning:

Time Investment: 1 hour