AI in Industry and the Physical World

AI in the physical economy is the use of machine learning in systems that sense and act in the physical world: industrial process control, robots in factories and warehouses, and vehicles that drive themselves. Safety here is judged by physical outcomes (injuries, crashes, energy use), so the distinction between a deployed product, a validated result and a research demonstration carries more weight than in software-only domains.

Three facts frame the field. First, most robots in factories are not learning systems: industrial robots are largely programmed manipulators, and the count of robots is not a count of AI systems. Second, the most mature learned-control deployments are narrow and heavily engineered, such as driverless ride-hailing in mapped service areas and control of building cooling plant. Third, general-purpose learned robot policies are, as of this draft, mostly research demonstrations. This article assigns each example to one of three classes:

Manufacturing, logistics and autonomous driving are treated together because they share the problem of reliable control under distribution shift; see TECH-ROBOT-001 for the underlying methods and TECH-RL-001 for reinforcement learning.

Industrial robots entered factories decades ago as programmed machines. Learned perception and control developed with deep learning and reinforcement learning (TECH-RL-001). Autonomous-vehicle development has a longer record of public testing; the first widely reported pedestrian fatality involving a vehicle under developmental automated-driving control occurred in Tempe, Arizona, on 2018-03-18, and the US National Transportation Safety Board (NTSB) investigated it (NTSB, 2019). In 2023 California suspended Cruise's driverless permits (TechCrunch, 2023). From 2023-25 academic robotics moved toward large pretrained "vision-language-action" models that map camera images and instructions to robot actions (Brohan et al., 2023; Black et al., 2024).

Levels of driving automation. The SAE J3016 taxonomy (April 2021 revision) defines six levels from 0 (no automation) to 5 (full automation). Levels 1-2 are driver-support features in which the human drives and supervises; Levels 3-5 are automated driving systems, and the level that applies depends on which feature is engaged, not on the vehicle's hardware (SAE International, 2021). Most robotaxi services operate at Level 4: the system drives without a human in the driver's seat within a defined operating domain. Claims about "self-driving" should always state the level.

Statistical difficulty of proving safety. Fatal crashes are rare per mile. Kalra and Paddock (2016) analysed how many miles of driving would be needed to demonstrate autonomous-vehicle reliability statistically and concluded that the required distances are very large [VERIFICATION NEEDED: specific mileage figures not read from the paper]. For a failure rate $\lambda$ per mile, observing zero events in $m$ miles supports an upper bound of roughly $3/m$ at 95% confidence (the "rule of three"), so demonstrating a rate below a human fatality benchmark requires correspondingly large $m$. Comparisons therefore often use more frequent outcomes (police-reported crashes, injury crashes, airbag deployments), which are statistical proxies.

Learned robot policies. A vision-language-action model maps an image $o_t$ and instruction $\ell$ to an action $a_t$: $a_t \sim \pi_\theta(a \mid o_t, \ell)$. RT-2 represented actions as text tokens and co-fine-tuned a vision-language model on robot trajectories and web data, evaluated in about 6,000 trials (Brohan et al., 2023). The $\pi_0$ model combined a pretrained vision-language model with a flow-matching action head trained on data from multiple dexterous robot platforms, and was evaluated on tasks such as laundry folding, table cleaning and box assembly (Black et al., 2024).

Reinforcement learning in industrial control. Controllers can be trained offline and in simulation, then constrained and tested live. DeepMind and Trane reported live experiments on two commercial cooling facilities, describing challenges in evaluation, learning from offline data and constraint satisfaction (Luo et al., 2022).

• Industrial automation: conventional programmed robots, machine-vision inspection, predictive maintenance, process optimisation.
• Warehouse and logistics robots: mobile and picking robots, mixing classical planning with learned perception.
• Autonomous vehicles: driver-support features (SAE Levels 1-2) versus driverless services (Level 4).
• Learned generalist robot policies: vision-language-action models for manipulation (research).
• Process and energy control: reinforcement-learning and model-predictive controllers for plants and buildings.

#6.1 Deployed products

• Industrial robots. The International Federation of Robotics (IFR) reported an operational stock of about 4.66 million industrial robots worldwide at the end of 2024 (up 9% year on year) and 542,000 installations in 2024, with Asia accounting for 74% of new installations (IFR, 2025, press release and executive summary, World Robotics 2025). These counts cover industrial robots broadly; they do not measure how many use machine learning.
• Driverless ride-hailing. Waymo operates a commercial Level 4 service. Press coverage in 2026 reports operation in ten US cities and about 500,000 paid rides per week (Electrek, 2026; Tier 4) [VERIFICATION NEEDED: ride counts and city count are company-reported and secondary-sourced; for current figures see the Living Reference collections]. Cruise's driverless permits in California were suspended by the state DMV in October 2023, which the DMV tied to safety risk and to the company withholding video from an investigation (TechCrunch, 2023); the programme subsequently lost its parent's funding (reported by press; date not verified here).

#6.2 Validated results

• Crash-rate comparisons for a Level 4 service. Analyses by Waymo-affiliated researchers, using federal crash reports and company-supplied mileage, compared the service with human benchmarks matched to vehicle type, road type and location. At 56.7 million rider-only miles through January 2025, the rider-only crashed-vehicle rate was statistically significantly lower than the benchmarks for any-injury-reported, airbag-deployment and suspected-serious-injury-or-worse crashes, with no statistically significant disbenefit found in any of 11 crash-type groups (Kusano et al., 2025; earlier analysis at 7.1 million miles: Kusano et al., 2024). Caveats: the authors are affiliated with the operator; mileage is company-provided; benchmarks have estimation error; the results apply to the operator's domains and software versions at the time. Independent replication was not verified in this draft.
• Accident investigation, developmental system. The NTSB found that the probable cause of the 2018 Tempe crash was the failure of the vehicle operator to monitor the driving environment and the automated system because she was visually distracted by a personal phone; the Board also cited the developer's safety culture and limited state oversight of testing (NTSB, 2019). The case shows failure at the interface between automation and human supervision, not only in perception.
• Industrial control. In live experiments at two commercial sites, reinforcement-learning control of cooling plant produced energy savings of approximately 9% and 13% (Luo et al., 2022). The paper is a technical overview from the developers, and two sites do not establish general performance.

#6.3 Research demonstrations

RT-2 and $\pi_0$ show that pretrained models can transfer semantic knowledge to robot control and handle varied manipulation tasks (Brohan et al., 2023; Black et al., 2024). Both are developer reports evaluated in laboratory-style settings; reliability and cost for continuous factory operation were not shown by these papers.

Advantages. Consistency and 24-hour operation; possible safety gains in driving where crash exposure is large (Kusano et al., 2025); energy savings in tuned control problems (Luo et al., 2022).

Limitations.

• Distribution shift and the long tail. Rare scenarios dominate safety risk and are hard to cover in training or tests.
• Human-automation interaction. Supervised partial automation has failure modes distinct from full automation (NTSB, 2019).
• Evidence quality. Much safety evidence comes from operators; benchmarks are imperfect; independent audits are limited.
• Domain-bound results. Findings for one operator, city or plant do not transfer automatically.
• Operational and regulatory fragility. A single incident and an adverse regulatory response ended a driverless programme's operations in a major market (TechCrunch, 2023).

Unresolved questions. How to certify learned controllers; how performance holds as service areas, weather and traffic complexity grow; whether generalist robot policies reach industrial reliability; effects on employment in logistics and driving (see ECON-LABOR-001).

As of 2026-10-15: Level 4 ride-hailing is a commercial reality in several US cities, with peer-reviewed operator-affiliated evidence of lower crash rates than human benchmarks (Well Supported with the caveats above; independent replication not verified). Industrial robot deployment is large and mostly conventional (Established as to counts, per IFR). General-purpose learned robot policies remain research demonstrations (Preliminary). Learned industrial control has working field results at small scale (Preliminary). Accident investigations and permit actions show that safety and governance failures can stop a programme.

Near-term (0-3 years): geographic and operating-domain expansion of driverless services; more independent safety analyses; pilot deployments of learned manipulation in structured settings. Medium-term (3-10 years): wider use of learned policies in warehouses and light manufacturing, conditional on reliability. Long-term claims about general-purpose humanoid labour are speculative and not rated.

• Foundational Concepts:
  • TECH-ROBOT-001 Robotics and Embodied AI: Methods for robot learning.
  • TECH-RL-001 Reinforcement Learning: Control learning used in industrial and robot settings.
• Related Methods & Technologies:
  • TECH-HW-001 AI Hardware: Onboard compute for vehicles and robots.
  • CAP-LIMITS-001 Hallucination, Reliability and Calibration: Reliability concepts relevant to safety cases.
  • APP-GOV-001 AI in Government and Public Services: Public regulation and oversight of automated systems.
• Advanced Implications & Future Trajectories:
  • GOV-LAW-001 AI Governance: Regulatory frameworks.
  • ECON-LABOR-001 AI, Productivity and Labor Markets: Employment effects.