Safe Autonomy: How AI‑Driven Robots Are Rewriting Industrial Robotics Engineering

For four decades, industrial robotics rested on a comfortable engineering assumption: a robot's behaviour is fully specified before it is switched on. Programs were deterministic, motion envelopes were fixed, and safety was largely a geometry problem — keep people and machine trajectories separated in space and time, and prove it.

That assumption is now under pressure from two directions simultaneously. On the technical side, learned policies — models trained on demonstration and simulation data rather than hand-coded waypoints — are moving from laboratory demonstrations into limited production use. On the regulatory side, the rulebook has been rewritten: the flagship industrial robot safety standard was reissued in 2025 after roughly eight years of revision work, a dedicated standard for balancing and legged robots is still in committee draft as of mid‑2026, and the European Union's new Machinery Regulation becomes applicable on 20 January 2027.

This article is written for the people who have to reconcile those two forces: automation and controls engineers, integrators, robotics researchers moving toward deployment, EHS and compliance leads, and the operations executives who sign off on capital projects. It sets out what has genuinely changed, what remains experimental, where the evidence is thin, and what a defensible engineering approach looks like today.

---

Executive summary

• Global industrial robot deployment remains at historically high volumes: 542,000 units were installed in 2024, bringing the worldwide operational stock to 4,664,000 units — a 9% year‑on‑year increase, with Asia accounting for 74% of new deployments, Europe 16% and the Americas 9%. International Federation of Robotics + 2
• The ISO 10218 series was technically revised and reissued in 2025, replacing the 2011 editions and absorbing the collaborative‑application requirements previously held in ISO/TS 15066. ISO
• Safety governance is shifting from device certification toward application‑level risk assessment — a change that matters most for robots whose behaviour is partly learned.
• There is no published international standard specific to humanoid or dynamically balancing industrial robots. ISO/CD 25785‑1 is at committee draft stage, with its comment period closing on 8 July 2026. ISO
• Robot foundation models (vision‑language‑action policies) have demonstrated meaningful cross‑task generalisation in research settings, but published robustness analyses show performance degrading under perturbations that industrial environments produce routinely.
• Documented commercial humanoid work is real but narrow: company‑reported figures describe tote handling and material transfer, not high‑precision assembly.
• The EU Machinery Regulation applies from 20 January 2027 and introduces explicit provisions for machine‑learning‑based safety components — a compliance deadline that should already be shaping 2026 design decisions.
• The binding constraint on adoption is increasingly verification and evidence, not raw capability.

---

The scale and shape of the current market

Robotics is not an emerging market; it is a mature capital‑goods sector with a fast‑moving software layer bolted on top. The International Federation of Robotics reported 2024 as the second‑highest installation year on record, roughly 2% below the 2022 peak. China alone installed 295,000 units in 2024 and passed two million units of operational stock, while domestic Chinese suppliers took 57% of their home market, up from 47% the previous year. The IFR's forecast anticipates about 575,000 installations in 2025 and more than 700,000 by 2028. International Federation of Robotics + 2

Two structural signals matter for engineers. First, the supplier base is regionalising — component sourcing, spare‑parts strategy and controller ecosystems are now geopolitical variables, not just procurement ones. Second, the IFR placed the market value of industrial robot installations at an all‑time high of US$16.7 billion while unit growth was flat, which suggests value is migrating toward software, integration and services rather than arms. Modern Materials Handling

Table 1 — Where value and risk are moving

Assessment by OneWise based on the sources cited in this article.

---

What actually changed in the standards

ISO 10218:2025 — the flagship revision

The third edition of ISO 10218‑1 was prepared by ISO/TC 299 (Robotics) in collaboration with CEN/TC 310 under the Vienna Agreement, and cancels and replaces the 2011 second edition. Part 1 addresses robot manufacturers; Part 2 addresses integrators, covering robot applications and robot cells. The revision followed roughly eight years of committee work and, according to A3, focuses on making functional safety requirements explicit rather than implied. ISO + 2

Three changes have the largest practical effect:

1. Collaborative operation is no longer a separate technical specification. Requirements for robots intended for collaborative applications — formerly the content of ISO/TS 15066 — are incorporated into the revised standard. The practical consequence is the retirement of a widespread misconception: there is no such thing as a "collaborative robot" that is inherently safe. Safety is a property of the application — robot, end effector, payload, workpiece, speed, and human task combined. ISO

2. Robot classification. Suppliers and standards commentators report that Part 1 introduces a classification distinguishing robot types, based on parameters including mass per manipulator, maximum force and maximum speed, reflecting the practical difference between large heavy manipulators and smaller robots intended for collaborative work. Engineers should verify the exact criteria against the purchased standard text rather than secondary summaries. IDEC APACSICK

3. Substantially expanded integration requirements. Part 2 grew to roughly three times its previous length, with the terms and definitions clause expanding from 2 to 15 pages and the risk‑reduction requirements from 28 to 50 pages, adding material such as risk assessment for contact between moving parts of the application and operators. The direction is unambiguous: more of the safety argument now sits with the integrator and the end user, not the robot vendor. IDEC APAC

The humanoid and legged‑robot gap

There is currently no published international Type C standard for balancing robots. ISO/CD 25785‑1 covers safety requirements for dynamically stable industrial mobile robots — legged, wheeled or otherwise — defining "actively controlled stability" as a robot that requires active control to remain balanced and could become unstable without power. The project was approved on 22 May 2025, the committee draft was registered on 8 May 2026, and the comment period closed on 8 July 2026. Exoskeletons, ridden robots, road vehicles and non‑industrial robots are explicitly out of scope, and a separate Part 2 addressing integration is planned. ISOISO

Until that work is published, deployments of legged and humanoid platforms in industrial settings rest on the general machinery framework, ISO 10218:2025 where applicable, manufacturer specifications and site‑specific risk assessment. That is a legitimate position — it is not a compliance vacuum — but it places the analytical burden on the deploying organisation, particularly for fall hazards, which have no direct analogue in fixed‑arm robotics.

Europe's 2027 deadline

Regulation (EU) 2023/1230 replaces Directive 2006/42/EC, has been in force since 19 July 2023, and applies in general from 20 January 2027. It introduces, for the first time, express provisions on systems with self‑evolving behaviour based on machine learning and on safety‑relevant cybersecurity. Machinery placed on the EU market before that date must comply with the 2006 Directive. Some provisions, such as those for notified bodies, applied earlier — from 20 January 2024. Ai-resources + 2

For a project with an 18‑month design‑to‑commissioning cycle, this is not a future issue. Products intended for the EU market in 2027 are being architected now.

Table 2 — Standards and regulation snapshot (status as of 1 August 2026)

Compiled by OneWise from ISO and EU sources cited in the references.

---

The technical shift: learned policies enter the plant

What robot foundation models are

The research direction that has dominated robot learning since 2024 is the vision‑language‑action (VLA) model: a policy built on a pretrained vision‑language backbone, fine‑tuned on robot demonstration data, that maps camera images and a natural‑language instruction directly to actions. Physical Intelligence's π₀ argued that generalist robot policies can address the data, generalisation and robustness obstacles in robot learning, using a flow‑matching architecture on top of a pretrained vision‑language model to inherit internet‑scale semantic knowledge, trained in distinct pre‑training and post‑training phases analogous to large language model practice. Google DeepMind's Gemini Robotics line and NVIDIA's GR00T N1 pursue comparable objectives, and the Open X‑Embodiment and DROID datasets supply the cross‑platform training data. arXivarXiv

The engineering appeal is obvious: less fixturing, faster changeover, tolerance of variation. The engineering problem is equally obvious.

Why this breaks conventional verification

A deterministic program can be exhaustively reviewed. A learned policy cannot. Published work from within the field is candid about this — a Google DeepMind report on evaluating Gemini Robotics policies in a video world simulator notes the goal of predicting how policies degrade along different generalisation axes such as scene objects and visual background, and describes the work as early‑stage. Recent robustness literature (for example, LIBERO‑plus and analyses of multi‑modal perturbation) reports that VLA performance can fall substantially under changes that a factory produces routinely — lighting shifts, camera pose drift, distractor objects, instruction rephrasing. arXiv

The correct engineering response is not to reject learned policies. It is to refuse to let them carry safety functions.

Table 3 — Deterministic control versus learned policy: engineering implications

Original framework developed for OneWise.

The architectural principle that follows is separation: let the learned policy propose, and let a deterministic, independently verified safety layer dispose. Speed and separation monitoring, safe torque off, geometric limiting and emergency stop functions belong in certified hardware and rated safety logic. The model operates inside an envelope it cannot widen.

---

Where autonomy is actually working

Evidence quality varies sharply here, and professionals should treat vendor announcements as claims rather than findings.

Documented, company‑reported: Agility Robotics' Digit has been deployed under a robots‑as‑a‑service agreement at a GXO Logistics facility in Flowery Branch, Georgia; the company reported in November 2025 that Digit had moved more than 100,000 totes in commercial operation, handling tasks including moving totes on and off autonomous mobile robots, loading conveyors and stacking containers within a live fulfilment workflow. The same programme, described in an NVIDIA case study, uses simulation‑trained controllers validated across a second physics engine before hardware deployment. These are meaningful operational data points; they are also single‑source and originate with commercially interested parties. Robotics & Automation NewsNVIDIA

Well established: high‑volume, high‑precision fixed‑arm work — welding, dispensing, machine tending, palletising — remains the domain of conventional industrial robots with deterministic control. Nothing in recent AI progress changes that.

Genuinely uncertain: whether generalist policies will reduce integration cost enough to open the small‑and‑medium‑enterprise market, and on what timescale. Forecasts here are forecasts, including the IFR's.

Table 4 — Maturity assessment

---

Figure descriptions for the OneWise design team

Figure 1 — "Layered Autonomy Architecture." Horizontal band diagram, four stacked layers. Top: Task intent (operator instruction, work order). Second: Learned policy (VLA model; perception → action proposal), shaded to indicate probabilistic behaviour. Third: Deterministic supervisor (envelope check, speed and separation monitoring, force limits), shown in a solid contrasting colour to signal certified logic. Bottom: Rated safety hardware (safe torque off, e‑stop, safety‑rated encoders). Downward arrows from top to bottom labelled "proposed action"; an upward arrow from the supervisor labelled "veto / safe state". A side annotation reads: Only the lower two layers carry safety functions. Caption: Separation of learned behaviour from verified safety functions.

Figure 2 — "Compliance Timeline 2011–2027." Horizontal timeline. Markers: 2011 (ISO 10218 second edition), 19 July 2023 (EU Machinery Regulation enters into force), 2025 (ISO 10218‑1/‑2 third edition published; ISO 25785‑1 project approved 22 May), 8 May 2026 (ISO/CD 25785‑1 registered), 8 July 2026 (CD comment period closes), 20 January 2027 (Machinery Regulation applies). Use a dashed segment after mid‑2026 to indicate that ISO 25785‑1 publication timing is not fixed. Caption: Standards and regulation moved faster in three years than in the preceding decade.

Figure 3 — "Risk Assessment Loop for Adaptive Systems." Circular diagram, five nodes: Hazard identification → Risk estimation → Risk reduction measures → Validation and evidence → Operational monitoring, with a return arrow from monitoring to hazard identification labelled "model update / substantial modification". Caption: For adaptive systems, risk assessment is a loop, not a one‑time deliverable.

---

Frequently misunderstood concepts

"Cobots are safe by design." No. Collaborative requirements now sit inside ISO 10218‑1:2025, and safety derives from the application — including the end effector and workpiece, which are frequently the actual hazard. ISO

"AI makes the robot safer because it can see people." Perception improves situational performance; it does not by itself deliver a safety function unless implemented in rated hardware with a determinable performance level.

"A software update is not a modification." Under the EU framework, substantial modification is a defined concept, and the Machinery Regulation contains express provisions for systems with self‑evolving behaviour based on machine learning. Model updates need change control equivalent to mechanical changes. Ai-resources

"Humanoids will replace industrial arms." The documented commercial work is material handling in human‑shaped environments — a complement to fixed automation, not a substitute for it.

---

Practical takeaways

1. Buy the standards, not the summaries. Secondary commentary on ISO 10218:2025 — including this article — is orientation. Design decisions require the normative text.
2. Rewrite the risk assessment template for adaptive behaviour. Add explicit sections for model provenance, evaluation coverage, out‑of‑distribution monitoring and update governance.
3. Architect for separation now. If a learned component can influence motion, the safety envelope must be enforced independently and be independently verifiable.
4. Treat 20 January 2027 as a design constraint. For EU‑market products, the conformity route — including whether an AI‑based safety component triggers third‑party assessment — should be settled before detailed design freezes.
5. Instrument deployments for evidence. Cycle counts, intervention rates, near‑miss logs and failure taxonomies are the only credible basis for scaling decisions, and they will also be the basis of any liability defence.
6. Pilot humanoids on defensible tasks. Tote and container movement with clear fall zones is a reasonable first case; anything requiring sub‑millimetre repeatability is not.
7. Budget for cybersecurity as a safety cost. IT/OT convergence — one of the IFR's identified 2026 trends — means connectivity assumptions now sit inside the safety case. The Robot Report
8. Plan the workforce transition explicitly. Skills for supervising, evaluating and troubleshooting learned systems differ from those for teach‑pendant programming.

---

Key insights

1. Installation volumes are high and roughly flat; the growth is in software, services and integration.
2. Safety responsibility has shifted decisively toward the integrator and end user.
3. Collaborative operation is an application property, permanently — not a robot category.
4. The humanoid safety standard is real and progressing, but not published; deployments proceed on general machinery law plus site risk assessment.
5. Learned policies generalise impressively and degrade quietly; both facts are load‑bearing.
6. No accepted method yet exists for certifying a learned policy as a safety function.
7. Regulatory attention has moved from mechanical hazards to behaviour, adaptation and cybersecurity.
8. Company‑reported deployment milestones are useful evidence but not independent verification.
9. The 2027 EU deadline is already inside most industrial design cycles.
10. The organisations that scale autonomy first will be those with the best measurement discipline, not the newest models.

---

Latest developments

• 8 January 2026 — The IFR published its Top 5 Global Robotics Trends for 2026, identifying agentic AI as a key trend combining analytical AI for structured decision‑making with generative AI for adaptability, alongside accelerating demand for versatile robots; the list also covered IT/OT convergence, humanoids, safety and security concerns, and labour shortages. This is an industry association's outlook, not peer‑reviewed analysis. International Federation of RoboticsThe Robot Report
• 8 May 2026 — ISO/CD 25785‑1 registered as a committee draft. ISO
• 8 July 2026 — Comment period on ISO/CD 25785‑1 closed (stage 30.60). Publication timing remains unset; any date circulating publicly is an estimate. ISO
• Ongoing — Commercial humanoid deployments continue to expand across logistics and manufacturing sites according to supplier statements. Independent, peer‑reviewed evaluation of throughput, safety incidents and total cost of ownership remains largely absent from the public record — the single most important evidence gap in this field.

---

FAQ

1. Is ISO 10218:2025 mandatory?
ISO standards are voluntary in themselves. They acquire force through national adoption, regulatory harmonisation and contractual or insurance requirements — in practice, compliance is close to obligatory for most industrial suppliers.

2. Does ISO/TS 15066 still apply?
Its collaborative‑application content has been incorporated into ISO 10218‑1:2025. New designs should work from the revised standard. ISO

3. Do robots installed before 2025 need to be re‑certified?
ISO 10218‑1:2025 states that it does not apply to robots manufactured before its publication date. Existing installations are generally governed by the framework in force at the time, though substantial modification can change that. ISO

4. What is a "dynamically stable" robot?
One requiring active control to remain balanced, which could become unstable without power — including bipedal, quadrupedal and wheeled balancing platforms. ISO

5. Can a learned model perform a safety function?
There is no established verification and validation method that would support it. Safety functions should remain in deterministic, rated implementations.

6. When does the EU Machinery Regulation apply?
From 20 January 2027, replacing Directive 2006/42/EC. Ai-resources

7. Does a model update count as a substantial modification?
Potentially, depending on whether it changes the machine's performance or intended use in a way not foreseen in the original risk assessment. Treat model updates under formal change control.

8. How many industrial robots are in operation worldwide?
4,664,000 units as of 2024, per the IFR's World Robotics 2025 report. International Federation of Robotics

9. Are humanoid robots ready for production?
For narrow material‑handling tasks, company‑reported deployments exist. For precision assembly, no.

10. What is a vision‑language‑action model?
A policy that maps camera input and a language instruction to robot actions, typically built on a pretrained vision‑language backbone.

11. Do foundation models eliminate robot programming?
No. They shift effort from waypoint programming toward data curation, evaluation and monitoring.

12. What is the main technical risk of learned policies in industry?
Out‑of‑distribution degradation — competent‑looking behaviour that fails under conditions absent from training data.

13. Which region leads robot adoption?
Asia, with 74% of 2024 installations; China alone accounted for 54% of global deployments. LinkedIn

14. Is simulation training trustworthy?
It reduces cost and hardware risk but introduces sim‑to‑real gaps. Some developers cross‑validate policies across more than one physics engine before deployment. NVIDIA

15. What should a first AI‑robotics pilot measure?
Intervention rate, cycle time variance, failure taxonomy and near‑misses — not demonstration success rate.

16. Does cybersecurity now fall inside the safety case?
Increasingly yes; the EU Machinery Regulation introduces express provisions on safety‑relevant cybersecurity. Ai-resources

---

Glossary

• Cobot — colloquial term for a robot designed for collaborative applications; not a formal safety classification.
• CD (Committee Draft) — an ISO development stage in which a draft is reviewed by the technical committee.
• Dynamically stable robot — a robot requiring active control to maintain balance.
• Functional safety — safety achieved through correct functioning of a safety‑related control system.
• IT/OT convergence — integration of information technology systems with operational technology on the plant floor.
• Partly completed machinery — equipment intended for incorporation into other machinery; the regulatory category covering industrial robot arms.
• RaaS (Robots‑as‑a‑Service) — a commercial model in which robots are provided under a service contract rather than purchased outright.
• Safe torque off (STO) — a rated safety function removing power capable of generating torque.
• Speed and separation monitoring — a risk‑reduction method maintaining a protective distance between human and robot.
• Substantial modification — a change to machinery significant enough to trigger renewed conformity obligations.
• VLA (vision‑language‑action) model — a learned policy mapping visual input and language instruction to robot actions.

---

References

Standards and standards bodies

International Organization for Standardization. (2025). ISO 10218‑1:2025 — Robotics: Safety requirements — Part 1: Industrial robots (3rd ed.). ISO. https://www.iso.org/standard/73933.html

International Organization for Standardization. (2025). ISO 10218‑2:2025 — Robotics: Safety requirements — Part 2: Industrial robot applications and robot cells. ISO. https://www.iso.org/standard/73934.html

International Organization for Standardization. (2026). ISO/CD 25785‑1 — Robotics: Safety requirements for dynamically stable industrial mobile robots (legged, wheeled, or other forms of locomotion) — Part 1: Robots (Committee draft, stage 30.60). ISO/TC 299. https://www.iso.org/standard/91469.html

Government and regulatory sources

European Union. (2023). Regulation (EU) 2023/1230 of the European Parliament and of the Council of 14 June 2023 on machinery. Official Journal of the European Union, 29 June 2023.

European Agency for Safety and Health at Work. (n.d.). Regulation 2023/1230/EU — machinery. EU‑OSHA. https://osha.europa.eu/en/legislation/directive/regulation-20231230eu-machinery

Industry reports and association publications

International Federation of Robotics. (2025, September 25). World Robotics 2025 report: Global robot demand in factories doubles over 10 years [Press release]. https://ifr.org/ifr-press-releases/news/global-robot-demand-in-factories-doubles-over-10-years

International Federation of Robotics. (2026, January 8). Top 5 global robotics trends 2026 [Press release]. https://ifr.org/ifr-press-releases/news/top-5-global-robotics-trends-2026

Association for Advancing Automation. (2025, February). Updated ISO 10218: Major advancements in industrial robot safety standards now available. https://www.automate.org/robotics/news/updated-iso-10218-major-advancements-in-industrial-robot-safety-standards-now-available

IDEC Corporation. (2026, March 3). Behind the ISO 10218 series safety standards updates in 2025 (Article 1 of 2). https://www.idec.com/en-apac/blog/iso-10218-updates-revisions-background-part-1

Academic and research papers

Black, K., Brown, N., Driess, D., Esmail, A., Equi, M., Finn, C., … Zhilinsky, U. (2024). π₀: A vision‑language‑action flow model for general robot control (arXiv:2410.24164). arXiv. https://arxiv.org/abs/2410.24164

Gemini Robotics Team. (2025). Gemini Robotics: Bringing AI into the physical world (arXiv:2503.20020). arXiv. https://arxiv.org/abs/2503.20020

Gemini Robotics Team. (2025). Gemini Robotics 1.5: Pushing the frontier of generalist robots with advanced embodied reasoning, thinking, and motion transfer (arXiv:2510.03342). arXiv. https://arxiv.org/abs/2510.03342

Bjorck, J., Castañeda, F., Cherniadev, N., Da, X., Ding, R., Fan, L., … Huang, S. (2025). GR00T N1: An open foundation model for generalist humanoid robots (arXiv:2503.14734). arXiv. https://arxiv.org/abs/2503.14734

Physical Intelligence. (2025). π₀.₅: A vision‑language‑action model with open‑world generalization (arXiv:2504.16054). arXiv. https://arxiv.org/abs/2504.16054

Khazatsky, A., Pertsch, K., et al. (2024). DROID: A large‑scale in‑the‑wild robot manipulation dataset. Robotics: Science and Systems.

Open X‑Embodiment Collaboration. (2024). Open X‑Embodiment: Robotic learning datasets and RT‑X models. IEEE International Conference on Robotics and Automation (ICRA).

Fei, S., Wang, S., Shi, J., et al. (2025). LIBERO‑plus: In‑depth robustness analysis of vision‑language‑action models (arXiv:2510.13626). arXiv. https://arxiv.org/abs/2510.13626

Guo, J., Wu, Z., Tu, C., et al. (2025). On robustness of vision‑language‑action model against multi‑modal perturbations (arXiv:2510.00037). arXiv. https://arxiv.org/abs/2510.00037

Company and deployment sources (interested parties; treat as claims)

Agility Robotics. (2025, November 20). Digit moves over 100,000 totes in commercial deployment. https://www.agilityrobotics.com/content/digit-moves-over-100k-totes

NVIDIA. (n.d.). Agility Robotics: Humanoid robot Digit — whole‑body control foundation [Case study]. https://www.nvidia.com/en-us/case-studies/agility-robotics-digit-humanoid-robot/

---

Editorial note: This article was researched from the sources listed above. Where evidence originates with commercially interested parties, that provenance is stated in the text. Standards content is summarised for orientation only; engineering decisions should be based on the purchased normative texts.