Humanoid Robots: Why Billions Are Flowing In Now

For decades, humanoid robots were the symbol of promises about the future that never arrived – expensive trade-show attractions that climbed a staircase and then got switched off. The turnaround came quietly, and it’s now backed by numbers: in the first half of 2026 alone, roughly $18.8 billion flowed into robotics startups, and the first humanoids have real track records – at BMW, for instance, Figure robots supported the production of 30,000 vehicles and handled over 90,000 components with millimetre precision. What happened?

Key takeaways

• AI advances in motion control, not mechanics, are the real breakthrough – robots today learn tasks by demonstration rather than laborious programming.
• Falling hardware costs from the EV supply chain (batteries, motors, sensors) make humanoids economically viable for the first time.
• Deployment follows a clear staircase: structured factory work first, then logistics and retail, and household and service last of all.
• Nearly 90 percent of global production currently sits with Chinese manufacturers – the technology has a geopolitical dimension.
• The most reliable test isn’t the demo video – it’s paid working hours and unit numbers from real factories.

The real breakthrough is called AI

The mechanics – motors, joints, hands – were never the core problem; the brain was. Classic robots had to be programmed for every movement, which made them useless outside fixed sequences. The turning point came from the same direction as the language-model boom: foundation models for movement, trained on huge volumes of video and simulation data, let robots generalise tasks – demonstration instead of programming. A robot that learns today to grip one component can transfer that same movement pattern tomorrow to a similar component it has never seen before – classic, hard-programmed systems lacked this transfer capability entirely.

At the same time, hardware costs fell thanks to the EV supply chain (batteries, motors, sensors). Only this combination makes the promise credible that a robot sorting boxes today can take over a different station next year – and it explains why the big AI labs and carmakers got in at the same time: Tesla is pushing ahead with Optimus pilot production, Chinese manufacturers like UBTech are targeting deliveries in the five-figure range, and nearly 90 percent of global production currently sits with Chinese firms – the geopolitical dimension runs alongside.

The realistic deployment staircase

The path into everyday life follows a clear logic – from structured to chaotic, in four stages:

1. Stage 1 (now): pilot projects in factories and warehouses – repetitive, physically demanding tasks within a clear framework; this is where the cost calculation plays out against staff shortages and night shifts.
2. Stage 2 (over the next few years): logistics and retail in variable but controlled environments – shelving, picking, returns. The bridge to our article on warehouse automation: the humanoid as a flexible complement to AutoStore and similar systems, deployable exactly where rigid systems hit their limits.
3. Stage 3 (later half of the decade): public and semi-public services – reception, simple cleaning tasks, assistance in controlled settings like clinics or hotels.
4. Stage 4 (end of the decade): service and household – technically the hardest, economically the largest. This is where the field overlaps with specialised household robots, which already handle individual tasks today but, unlike the humanoid, aren’t general-purpose. The fair parallel for the humanoid remains autonomous driving: the curve is real, but longer than any keynote claims.

Each stage builds on the data volume of the one before it: a robot that has learned robust gripping and balancing across thousands of factory hours brings that foundational skill into the next, less structured environment. That’s exactly why the sequence isn’t a coincidence but a technical necessity – jumping straight to stage 4 without the training data of the earlier stages would be difficult to justify safely with today’s technology.

Worked example: when does a robot pay off?

A humanoid robot currently costs, depending on specification, roughly the high five figures to the low six figures – comparable to industrial equipment. Running it across multiple shifts, in a night shift that’s hard to staff, say, this investment can pay for itself over two to three years, provided it takes on tasks reliably enough to relieve a full position. That’s why the first economically sensible deployments emerge in factories with chronic staff shortages, not in well-staffed offices – there, the alternative to the robot isn’t a labour-cost problem, it’s simply: no staff available.

Who’s behind the billion-dollar race

The race splits across three recognisable camps, each bringing different strengths: US AI labs and Tesla bring foundation models and manufacturing experience from EV production. Established robotics specialists like Figure or Agility bring years of experience with real hardware and the first commercial pilot contracts. Chinese manufacturers like UBTech bet on scale effects from a supply chain that already accounts for most of global component production. These three camps compete not only for market share but also for standards: whose foundation model and whose hardware platform wins out will, in the long run, also decide technological dependencies – an issue that goes beyond pure economics.

Camp	Strength	Current focus
US AI labs & Tesla	Foundation models, manufacturing experience from EV production	Pilot production, internal factory deployments
Established robotics specialists	Years of hardware experience, first commercial contracts	Industrial and logistics pilots with third-party customers
Chinese manufacturers	Scale effects, dominant component supply chain	Series production, aggressive unit targets

From experience: anyone wanting to follow the industry seriously should pay less attention to launch events and more to quarterly reports from suppliers and pilot customers – that’s where unit numbers, contract values and actual operating hours show up, long before they appear in glossy presentations. Analyst reports on the robotics supply chain (actuators, sensors, battery cells) often provide a more honest progress indicator than any promo video.

The most common mistakes in assessing the field

1. Mistaking demo videos for market readiness. Fix: only unit numbers, contracts and paid working hours count as evidence, not choreographed clips.
2. Expecting household deployment within the next one to two years. Fix: the household is the last deployment stage, not the first – unstructured, safety-critical, expensive.
3. Lumping all manufacturers together. Fix: foundation-model quality, manufacturing experience and supply-chain access differ massively between providers.
4. Ignoring the geopolitical component. Fix: whoever dominates the component supply chain also influences price and availability in the long run – independent of pure software quality.
5. Confusing scepticism with rejection. Fix: the progress in motion AI is real and demonstrable – scepticism should target timelines and deployment readiness, not the underlying technology.

The sober observer’s position

Between “bubble” and “revolution”, one simple test helps: paid working hours instead of demo videos. Choreographed clips prove nothing; contracts, unit numbers and productivity data from real factories prove everything. Exactly this kind of evidence is starting to accumulate – slowly enough to warrant scepticism about the details, clearly enough to take the direction seriously.

The bottom line

Humanoid robots are no longer science fiction, but they aren’t ready for everyday series production either. The solid progress lies in structured industrial and logistics settings, not the living room – anyone wanting to assess the industry should look at unit numbers and employment contracts, not keynote choreography. The most interesting bet in the hardware world is under way, and the next sensible thing to watch is the quarterly figures from pilot projects, not the next product announcement.