ZharfAI Team
The robotics story that entered 2025 was dominated by humanoid demonstrations. The story visible by July 30, 2026 is more disciplined: industrial robots continue to scale, mobile and collaborative systems are broadening the set of automatable tasks, and humanoids are being tested in selected workflows—but a polished demo is not evidence of a reliable production system.
That distinction matters to an operator deciding what to buy. The useful unit of analysis is not “the robot” or the model behind it. It is a complete work cell: task, fixtures, sensors, controls, people, recovery procedures, safety case, maintenance plan, and evidence that the system meets a target under the real distribution of conditions.
The International Federation of Robotics recorded 542,076 industrial-robot installations in 2024, the second-highest annual total in its series, and an operational stock of about 4.66 million units. Electronics represented 24% of new installations and automotive 23%. Those figures describe measured 2024 activity reported in 2025; IFR numbers for later years are forecasts, not completed deployments.
The installed base is therefore large and growing, but it is concentrated in tasks that can be engineered: handling, welding, assembly, dispensing, inspection, palletizing, and machine tending. This is a better starting point than viral videos. It also explains why the shortage of capable integrators, especially for smaller manufacturers, can be a more immediate constraint than the availability of an impressive foundation model.
For a closer look at the factory context, see our analysis of AI in smart manufacturing. Robotics succeeds when it is integrated with production data and operating discipline, not when it is treated as a freestanding gadget.
A human-shaped machine can potentially use aisles, stairs, shelves, tools, and workstations designed around the human body. That is a genuine design advantage in brownfield facilities. It does not prove that a general-purpose humanoid is the lowest-risk or lowest-cost solution for a particular job.
Public demonstrations typically reveal little about intervention rate, task resets, teleoperation, fault recovery, battery swaps, payload variation, cycle-time distribution, safety separation, or performance after hundreds of hours. Until those measures are published under representative conditions, claims of “factory ready” should be read as a hypothesis to test.
Compare a humanoid with simpler alternatives: a fixed arm plus a fixture, an autonomous mobile robot, a conveyor change, a lift-assist device, or a redesigned tote. A specialized machine may be less flexible but easier to validate, repair, insure, and run at takt time. The winning form factor is the one that satisfies the work requirement and its hazards—not the one that looks most human.
A useful robotics architecture separates concerns:
AI can improve perception, instruction interpretation, planning, anomaly detection, and recovery suggestions. Safety-rated functions and hard motion limits remain separate engineering responsibilities. A language model should not be the sole component deciding whether a person can enter a hazardous space.
Our piece on autonomous robotics and human-machine collaboration explores this division of authority in more depth.
NIST’s robotics performance work organizes evaluation around capabilities such as perception, mobility, dexterity, and safety, while emphasizing the integrated system and its intended environment. A procurement trial should follow the same logic.
Build a representative test set of objects, locations, lighting, occlusions, packaging damage, floor transitions, human interruptions, and upstream mistakes. Measure task-success rate without help, human interventions per operating hour, safe-stop frequency, recovery time, cycle-time percentiles, damage, near misses, and availability. Average cycle time alone hides the long tail that disrupts production.
Test deliberately difficult but credible conditions. A gripper that handles 98% of ideal cartons may still fail economically if the remaining 2% creates a blocked conveyor every shift. Record why each failure occurred, whether the system detected it, and whether recovery made the situation safer or worse.
ISO 10218-1:2025 and Part 2 were revised in 2025. Part 1 addresses safety requirements for industrial robots and robot units; ISO 10218-2:2025 addresses integration, commissioning, operation, maintenance, and decommissioning of industrial robot applications and cells. The division is practical: the machine supplier and the cell integrator have related but distinct responsibilities.
Do not apply an industrial-robot standard blindly to every service, medical, consumer, or public-space robot; Part 1 explicitly identifies exclusions and domain boundaries. Likewise, the EU Machinery Regulation is a jurisdiction-specific legal instrument, not a universal global rule. Determine the applicable law, standards, site rules, and insurer requirements for each deployment.
OSHA notes that many robot accidents occur during non-routine work such as programming, maintenance, testing, setup, and adjustment. Risk assessment must therefore cover entry, clearing jams, teaching, tool changes, calibration, cleaning, and loss of power—not just the automatic cycle visible in a demo.
Robotics can remove repetitive lifting, exposure, and monotonous handling, but “lights-out” language obscures the human work required to keep a system trustworthy. Operators identify abnormal material, technicians restore degraded equipment, process engineers tune fixtures, safety staff investigate near misses, and supervisors decide when to stop.
Define authority before launch. Who may pause a cell? Who can approve a new object class, route, model, or grasp policy? When is manual recovery allowed? What information must a technician see before entering? A clear escalation path is a production feature.
Training should include failure recognition and safe recovery, not merely a happy-path user interface. Frontline observations should feed controlled improvement. Our guide to industrial copilots for frontline workers describes how software assistance can support that work without quietly transferring unsafe decisions to a model.
A robot must receive the correct job and material identity, respect quality holds, coordinate with people and other equipment, and write back a trustworthy result. That requires stable interfaces to warehouse, manufacturing-execution, maintenance, identity, and safety systems.
Design explicit states: requested, accepted, in progress, completed, rejected, paused, failed safely, and awaiting human review. Every command needs an idempotent identity so reconnects do not duplicate a move or transaction. Time synchronization, versioned recipes, device identity, network segmentation, and tamper-evident event history become important once the robot affects inventory or quality records.
Treat perception models and policies as controlled production components. Record version, configuration, sensor calibration, task context, decision confidence, and intervention. A model update that improves a benchmark can still degrade one supplier’s reflective packaging or a night-shift lighting condition.
The business case begins with a stable baseline: units per hour, labor and overtime, ergonomic exposure, scrap, damage, downtime, floor space, changeover, and the value of traceability. Then include fixtures, guarding, integration, validation, training, spares, connectivity, energy, software, cybersecurity, maintenance, and expected interventions.
Run sensitivity analysis for utilization, product mix, wage assumptions, downtime, and useful life. A robot that is excellent on one high-volume SKU may be idle after a mix change. Flexibility has value, but it must be demonstrated as reduced changeover effort or successful reuse—not assumed from the label “general purpose.”
Start with a bounded workflow whose exceptions are observable. Shadow the existing process, conduct a supervised pilot, compare against the baseline, and expand only after safety and operational gates pass. Preserve a rollback path and ownership for unresolved failures.
Before approving production, require evidence for the following:
Robotics is advancing, but adoption will remain uneven because physical work combines perception, contact, variability, and safety. The most credible deployments will often look less spectacular than a stage demonstration: constrained tasks, engineered environments, measurable human supervision, and careful expansion.
Humanoids deserve evaluation where existing human geometry creates a real advantage. They do not deserve an exemption from the evidence demanded of other capital equipment. The durable “robotics revolution” is a portfolio of well-integrated systems that make work safer and more reliable, with humans retaining clear authority over exceptions and change.
How artificial intelligence is pushing beyond Moore's Law, automating chip layout design, and accelerating the production of the advanced microchips powering the digital age.
Read More
Person-led elder care uses AI for bounded support with ongoing consent, accessible control, evidence, human response, and a safe option to refuse or stop.
Read More
Watchmaking AI can inspect parts and organize timing evidence, but calibrated tools, craft judgment, provenance, and reversible restoration govern the work.
Read MoreSee the daily briefing and the operational guides. This page is an archive note, not an invitation to start a project.