Humanoid Robots: From Factory Experiment to Everyday Technology

Robot demonstration videos are a poor evidence base. They are short, curated, usually staged in prepared environments, and silent about failure rates, uptime, energy draw and cost per hour of useful work. Yet they dominate public perception of where robotics stands.

A more useful way to assess humanoid robots is to ask what the form factor is actually for. The case is not that a human shape is optimal — for most single tasks it clearly is not. The case is that the world is already built for human bodies: door handles, stairs, shelf heights, tool grips, vehicle cabs, workbenches. A machine with roughly human proportions can, in principle, work in that environment without rebuilding it. That is the entire bet.

What robots do well today

Industrial robotics has been commercially successful for decades in a specific niche: high-volume, low-variation tasks in structured environments. Welding, painting, palletising, pick-and-place with known parts, and precision assembly of consistent components.

Recent progress has extended the envelope meaningfully:

• Mobile robots in warehouses. Autonomous transport of goods within mapped facilities is mature and widely deployed.
• Learned manipulation. Policies trained from demonstration and simulation now handle a wider range of objects than hand-coded control could.
• Better perception. Cheaper depth sensing and stronger vision models make cluttered scenes tractable in ways they were not a decade ago.
• Sim-to-real transfer. Training in simulation and deploying to hardware works well enough to accelerate development, though the reality gap persists.

Meanwhile, the International Federation of Robotics and national statistics show industrial robot installations concentrated heavily in a handful of manufacturing economies — a reminder that robotics adoption is currently an industrial-policy story as much as a technology one.

What remains genuinely hard

Dexterity in unstructured environments

Human hands perform continuous force modulation, incidental tactile sensing and rapid failure recovery. Robots handle deformable materials — cloth, cables, food, packaging — poorly, and struggle where the task cannot be fully specified in advance. Most household work falls in exactly this category, which is why the kitchen has been the field's polite embarrassment for thirty years.

Energy and duty cycle

A legged, arm-equipped machine spends energy on balance and motion whether or not it is doing useful work. Battery energy density improves slowly relative to compute. A robot that works for two hours and charges for one has a very different economic profile from one that works a full shift.

Reliability at scale

Commercial viability requires thousands of hours between meaningful failures, plus predictable maintenance. Novel hardware rarely achieves this early. Actuators wear, sensors drift, cables fatigue.

Safety and certification

Machines operating near people are governed by real standards — risk assessment, force and speed limits, separation monitoring, protective stops. Certification takes time and constrains design; it is not an obstacle to be disrupted away.

Cost

The relevant comparison is not the sticker price but the total cost per hour of useful output, including integration, supervision, maintenance, downtime and financing, measured against the wage and productivity of the alternative. In many low-wage tasks the arithmetic does not currently work.

Sector by sector

Manufacturing

The most likely early beachhead: semi-structured tasks in facilities that already understand automation, already have safety governance, and already measure cost per unit. Machine tending, kitting, moving parts between stations. The advantage of a general-purpose platform is redeployment — one machine covering several low-volume tasks instead of a dedicated cell per task.

Warehousing and logistics

High-volume, repetitive, measurable, and already heavily automated in transport. The remaining hard part is manipulation of varied items, particularly picking mixed goods and handling awkward packaging. Progress here is real and incremental.

Healthcare and elder care

Ageing populations create genuine demand, documented across WHO's work on ageing and OECD health data. But care is the hardest possible environment: unpredictable settings, fragile people, high liability, and a task content that is substantially relational rather than mechanical.

The realistic near-term contribution is logistical, not personal — moving supplies, disinfecting, delivering meals, lifting assistance under staff control, reducing the walking that consumes a large share of nursing time. Robots providing intimate personal care are a distant and ethically contested scenario, not a near-term product.

Construction

Attractive on paper: labour shortages, physically punishing work, measurable productivity problems. Difficult in practice: sites are unstructured, weather-exposed, constantly changing and organised around subcontracting. Expect narrow specialised machines — layout marking, drilling, bricklaying, rebar tying — before general humanoids.

Households

The most emotionally compelling market and the least likely to arrive soon. Homes are cluttered, unique, occupied by children and pets, and price-sensitive. A domestic robot must be cheap, safe, quiet, unsupervised and tolerant of chaos — a combination that no current platform approaches. Specialised appliances will continue to win.

The economics that actually decide it

Three variables dominate:

1. Cost per hour of useful work. Falls with hardware cost, rises with downtime and supervision. A machine needing frequent human intervention is a co-worker, not a replacement.
2. Task coverage per unit. A platform that performs one task competes with a cheaper dedicated machine. Its advantage must come from breadth.
3. Deployment friction. Integration, training, safety review and process change. Historically this exceeds hardware cost in industrial automation, and it is where most pilots quietly die.

Manufacturing scale can bring hardware costs down substantially — that is what happened with drones, batteries and solar. But scale requires demand, and demand requires reliability, which requires deployment. Bootstrapping that loop is the industry's central challenge.

Labour effects

Robotics affects work differently from software automation. Physical tasks are geographically fixed and often already short of workers. Where labour shortages are binding — logistics at peak, agriculture, some care logistics — robots substitute for vacancies rather than for people.

Where labour is plentiful and cheap, robots compete directly, and the effects are concentrated in specific regions and occupations. The distributional question is therefore local. National averages will conceal it, as the task-level analysis of AI and employment argues for cognitive work.

Acceptance and trust

Adoption depends on how people feel about machines in shared space. The evidence from existing deployments suggests acceptance is highest when the robot's behaviour is legible — predictable motion, clear signalling of intent, obvious stopping behaviour — and lowest when it is startling or ambiguous.

Humanoid form raises additional issues. Human-shaped machines invite attributions of intent and competence that the machine cannot support, and near-human appearance can provoke discomfort. In care settings there is a further question about substituting machine presence for human contact, which is a values question rather than an engineering one.

Three scenarios for 2035

Scenario A — Industrial normalisation (most plausible). Humanoid and semi-humanoid platforms find durable niches in manufacturing, logistics and inspection. Fleets number in the tens or low hundreds of thousands globally, concentrated in a few economies. Households remain served by specialised appliances.

Scenario B — Cost breakthrough. Actuator and battery costs fall sharply while learned control generalises better than expected. Robots enter service sectors — hospitality, retail, cleaning, some construction — and the labour effects become macroeconomically visible. This would require several favourable developments at once.

Scenario C — Prolonged pilot phase. Capital enthusiasm outruns unit economics; deployments stay small; consolidation follows disappointment. The technology continues improving quietly, as it did in the decades after earlier robotics hype cycles.

Indicators worth tracking

• Fleet operating hours and mean time between interventions, disclosed rather than claimed.
• Hardware bill-of-materials cost, particularly actuators and hands.
• Repeat orders from the same customer — the clearest signal that the economics work.
• Safety certifications for mobile manipulators working near people.
• Whether deployments replace vacancies or existing staff.

Conclusion

Humanoid robots are neither imminent domestic help nor a hoax. They are a plausible engineering bet on reusing human-shaped infrastructure, currently limited by hands, batteries, reliability and cost rather than by intelligence.

The place to watch is the loading bay and the factory floor, where the arithmetic is measurable and the environment is semi-structured. If the economics work there, the technology spreads outward. If they do not, no amount of impressive footage will change the outcome.