Humanoid Robots: What AI Assistants Can Really Do and What Comes Next

The Robot Is Leaving the Lab: What Humanoids Can Really Do — and What Still Stands Between Them and Our Homes

Humanoid robots are no longer just learning to walk. The real race is to make physical AI reliable, general, safe and economical enough to work in a world designed for humans.

Humanoid robots are beginning to move beyond laboratories and controlled factory tasks. The real test is whether embodied AI can become reliable enough to work safely in the unpredictable environments humans already inhabit.

The most important robot breakthrough is not walking

For decades, a humanoid robot was easy to recognize and hard to make useful. It could walk across a laboratory, climb stairs, dance, lift a box or perform a carefully rehearsed sequence. Those demonstrations solved genuine engineering problems, but they also encouraged a misleading idea: that once a robot could move like a person, a general-purpose assistant was almost here.

It was not. Walking is only one layer of the problem. A useful assistant must understand an instruction that was never written as code, perceive a room it has never seen, decide what matters, move without hurting anyone, grasp objects whose weight and friction it does not know, notice when something has gone wrong, recover, and keep doing useful work for hours. Then it must do all of that again tomorrow.

That is why the humanoid boom of 2026 is more interesting than the viral videos suggest. The International Federation of Robotics reports that roughly 7,000 full-size humanoids were sold globally in 2025, but it also stresses that most current applications remain specialized and often involve human teleoperation. In other words, the field has crossed the line from “almost nothing exists” to “an industry is forming,” but it has not crossed the line to general-purpose labor.

The useful way to think about the next few years is therefore not “When will we get C-3PO?” It is: what must a robot prove before we should treat it as a worker or a household assistant rather than an impressive machine?

The central idea of this article: the humanoid race is no longer mainly about building a human-shaped body. It is about turning AI into dependable physical behavior in environments that were never designed for robots.

1. Humanoid, android, robot assistant: these are not the same thing

A humanoid robot copies the basic geometry of the human body: torso, arms, usually legs, often hands and a sensor “head.” An android goes further and tries to look recognizably human, sometimes with a face, skin, hair or expressive features. A robot assistant is a functional category: it is defined by what it can do for people, not by how human it looks.

That distinction matters because the most serious commercial programs today are mostly building humanoids rather than realistic androids. Figure 03, Boston Dynamics Atlas, Apptronik Apollo, Agility Digit, Tesla Optimus, Unitree’s humanoids and 1X NEO are not trying to pass as human beings. The body plan is useful because our world is already standardized around us: handles, shelves, stairs, carts, tools, workbenches and kitchens are placed where human hands and feet can reach them.

This does not mean that the human body is the ideal machine. A fixed industrial arm is faster and more precise at repetitive assembly. Wheels are more energy-efficient than legs on a flat floor. A conveyor moves boxes better than a person-shaped robot. The business case for a humanoid begins only where the task is too varied to justify a dedicated machine, but the environment is too human-centered to redesign cheaply.

That is the first filter for the hype: a humanoid is valuable not because two legs look futuristic, but because one flexible machine might eventually use many spaces, tools and workflows that already exist.

2. The four thresholds between a demo and a real assistant

Humanoid robotics is unusually easy to overestimate because a short video hides the variable that matters most: repetition. A robot can succeed once and still be nowhere near deployable. To judge progress, it helps to separate four different thresholds.

Threshold	What it proves	What it does not prove	The metric that matters
1. Demonstration	The hardware and software can complete a task at least once.	Reliability, autonomy, economics, performance outside the prepared scene.	Was it autonomous? How many attempts and resets were needed?
2. Repeatable skill	The robot can perform a known workflow repeatedly in a known environment.	Transfer to other sites, new objects or unexpected conditions.	Success rate, intervention rate, cycle time.
3. Deployment	The robot creates useful output over long periods in a real operation.	General intelligence or household versatility.	Uptime, useful work-hours, maintenance, total cost per task.
4. General assistant	The robot transfers skills to unfamiliar environments and recovers from novel problems.	Human-level common sense or social understanding.	Zero-/few-shot transfer, recovery, safety, cost of supervision.

This framework changes how current announcements look. A foundation-model demonstration can be scientifically important without being commercially ready. A factory pilot can create real value without being “general intelligence.” A home robot can be useful while still depending on remote human assistance. These are different achievements, and collapsing them into a single word — autonomous — makes the entire field harder to understand.

3. Why the field changed so quickly: the robot finally got a more general brain

The hardware story is important but not sufficient. Electric actuators have improved, batteries are better, cameras and inertial sensors are cheap, compact computers are powerful, and tactile sensors are becoming practical. Simulation can now expose a control policy to millions of falls, collisions and variations before the first real robot moves.

The deeper change is software. Classical robots are excellent when engineers can define the world in advance. Detect this part. Move to these coordinates. Apply this force. Repeat. Human environments break that assumption because the state space explodes. A kitchen drawer is not always in the same place. A towel changes shape. A carton can be empty or full. A person can step into the path. A task described in one sentence may require twenty physical decisions.

That is where vision-language-action models, or VLAs, enter the picture. A 2026 IEEE Transactions on Robotics survey describes the core problem as the integration of perception, semantic understanding, task reasoning, physically grounded action generation and reliable execution. That final word matters: a robot is not useful merely because it “understands” the instruction. It has to make the world end up in the intended state.

Google DeepMind’s Gemini Robotics 2 is a good example of the new architecture. The system links high-level embodied reasoning with a VLA that controls physical action, including whole-body motion and dexterous manipulation. Figure’s Helix family follows the same broad direction: rather than programming a separate controller for every chore, the goal is a learned policy that maps perception and goals into coordinated behavior.

This is the robotics equivalent of the shift from specialized software to foundation models. The promise is not that one model is magically intelligent. The promise is transfer: experience gathered from many tasks, environments and robot bodies may make the next task cheaper to learn.

A useful robot needs more than computer vision. Modern embodied AI connects perception, reasoning, motor control and sensory feedback into a continuous loop — allowing the machine to see an object, plan what to do, act and correct itself when reality does not match the plan.

4. The real frontier is generalization, not choreography

A robot performing a known motion in a known room is a control problem. A robot succeeding in a room it has never seen is a generalization problem. The difference is enormous.

Figure made that distinction explicit with Helix 2.5, announced in September 2026. The company tested a policy across 30 unfamiliar homes and framed the project around zero-shot transfer: could the robot enter a new environment and immediately perform whole-body tasks without collecting training data in that specific home? Those results are company demonstrations, not an independent benchmark of household reliability, but the target itself is exactly right.

A useful home robot cannot require an engineer to remap every apartment. The owner should be able to move the coffee machine, buy different dishes, leave a bag on the floor and still expect the robot to cope. That is why “works in a second kitchen” is scientifically more meaningful than “performs a more spectacular trick in the first kitchen.”

Generalization also includes error recovery. People rarely execute a physical task exactly as planned. We reach for a cup, realize it is heavier than expected, shift our grip and continue. We open the wrong cupboard and immediately correct. We see that a chair blocks the route and alter the route. In robotics, each of those ordinary adjustments requires perception, a model of the task state and a policy capable of changing course.

A robot that never makes mistakes is unrealistic. A robot that notices and repairs its mistakes is a plausible product.

5. Where humanoids are actually working now: factories before kitchens

The first meaningful market is industrial because factories reduce uncertainty without eliminating the need for flexibility. Lighting is controlled, floors are predictable, work areas can be designed for safety, and the value of each task can be measured. At the same time, many factory jobs still involve moving parts, loading containers, sequencing components and using workspaces originally designed for human workers.

At BMW’s Spartanburg plant, Figure 03 is being used to develop a parts-sequencing workflow. Figure says the previous Figure 02 deployment contributed to production involving 30,000 vehicles; the new work is more complex because the robot must coordinate manipulation with whole-body repositioning and cart movement. The important point is not the vehicle count by itself — it is that the company is moving from a single manipulation station toward a longer logistics workflow.

Agility Robotics has taken an intentionally less cinematic route. Its Digit 5 was announced in September 2026 after, according to the company, more than 65,000 operating hours across earlier Digit systems. Agility emphasizes close operation around people, maintainability and facility workflows — exactly the unglamorous requirements that determine whether a robot survives outside a demo stage.

Boston Dynamics and Hyundai are following a similar commercialization path with Atlas. In September 2026, Boston Dynamics opened the Robotics Metaplant Application Center in Georgia, where Atlas is being trained on real automotive logistics such as parts sequencing and material handling. Hyundai plans to begin plant deployment in 2028 and expand toward assembly later if validation succeeds.

These examples reveal what the near-term humanoid economy is really about: not replacing every worker with a robot that can do everything, but finding enough adjacent tasks that one flexible machine can justify its cost.

6. The company race — viewed through what each program is trying to prove

Program	What it is trying to prove	Current evidence	Main unanswered question
Figure 03 / Helix	One learned stack can move from factory workflows toward general household tasks.	BMW logistics work; whole-body VLA demos; unfamiliar-home experiments.	Can reliability survive long deployments and new environments without frequent rescue?
Boston Dynamics Atlas	Exceptional mobility can become a maintainable industrial product.	Production-version hardware; RMAC training; committed 2026 deployments and Hyundai roadmap.	Can dynamic capability translate into competitive cost per useful task?
Agility Digit	A simpler humanoid can become ordinary logistics infrastructure.	Commercial deployments, large operating-hour base, Digit 5 safety focus.	How many workflows can one platform cover without expensive customization?
Apptronik Apollo 2	Large robot fleets can generate the data needed to train more general policies.	Robot Park fleets, bipedal and wheeled variants, DeepMind partnership.	Can the data flywheel produce transferable skills faster than deployment costs rise?
1X NEO	A home-first robot can become useful before full autonomy is solved.	Consumer-facing chores, self-charging, remote Expert Mode.	Can remote assistance shrink enough for privacy, economics and trust to work at scale?

Unitree	Cheaper humanoid bodies can accelerate research and ecosystem growth.	Commercial humanoid platforms at unusually low hardware prices.	How much capability and reliability can be delivered at those price points?
Tesla Optimus	Automotive-scale manufacturing can make humanoids a mass product.	Large investment and internal development; limited public evidence of broad autonomous deployment.	Can manufacturing scale be matched by mature dexterity, autonomy and uptime?

This comparison is more useful than a “best robot” ranking because the companies are attacking different bottlenecks. Some are trying to prove autonomy. Some are proving production. Some are proving safety. Some are proving that the hardware can become cheap. The eventual winner, if there is one, may be the company that solves the least glamorous combination: enough intelligence, enough reliability and a business model that survives maintenance.

There is no single model for the humanoid robot of the future. Some machines are being optimized for factories and logistics, others for dexterous manipulation, research or domestic assistance. The winning design may depend less on looking human than on solving a specific class of physical problems reliably.

7. Hands are the real nightmare

Walking attracts attention because it is visible. Manipulation is where useful work lives.

The human hand can lift a suitcase and then pick up a coin. It can feel a slipping glass before the eye notices it, deform a sponge without crushing what is inside, guide a cable through a gap and change grip while the rest of the body is moving. We do this with a dense combination of touch, proprioception, vision and learned physical intuition.

A 2026 IEEE/ASME survey of humanoid locomotion and manipulation highlights tactile sensing and whole-body feedback as increasingly important for contact-rich tasks. This is not a minor hardware upgrade. It changes what a robot knows about the world. Vision can estimate where a cup is; touch tells the robot whether the cup is actually secure in its fingers.

Soft objects make the problem worse. Laundry has no fixed geometry. A bag collapses when grasped. A cable bends, snags and disappears behind furniture. A human can infer these states from tiny cues. A robot must learn them from sensors and data, often while its own action changes the object it is trying to understand.

That is why apparently trivial demonstrations — tying a knot, loading a dishwasher, opening packaging, folding a shirt — matter. They are tests of contact, uncertainty and recovery. A humanoid that can walk beautifully but cannot handle the physical messiness of ordinary objects is a mobile camera stand, not an assistant.

8. The data problem: robots cannot learn from the internet alone

Language models became powerful partly because humanity had already created a vast digital training set: books, websites, code, images and video. Robotics does not have an equivalent archive of high-quality action data.

A video of a person making coffee is useful, but it does not directly contain joint torques, finger forces, exact camera geometry, tactile feedback or the sequence of corrective motions that kept the cup from slipping. Robot data is expensive because someone has to collect it with a physical system, often slowly.

This is why so many current programs are really data factories. Apptronik’s Robot Park runs fleets of Apollo 2 robots to gather real-world demonstrations. Figure is building its Index dataset around human behavior. NVIDIA says GR00T 1.7 was pretrained on roughly 32,000 hours of real data and 8,000 hours of simulated data. The numbers are not directly comparable, but the strategy is the same: collect enough diversity that the next skill does not start from zero.

Simulation helps because it is cheap and parallel, but a 2026 review of sim-to-real reinforcement learning emphasizes the persistent “reality gap”: small mismatches in dynamics, sensing, friction or environmental variability can break a policy that looked excellent in simulation.

This creates a powerful flywheel. More deployed robots generate more failures and successful trajectories. Those data improve models. Better models make deployment cheaper. Cheaper deployment creates more data. If humanoids scale, data may become as strategically important as motors and batteries.

9. The uncomfortable truth about autonomy: sometimes there is still a person behind the robot

Teleoperation is essential to modern robotics. A remote human can demonstrate a task, rescue a robot, collect training data and keep an early deployment productive while autonomy is incomplete. The mistake is not using teleoperation. The mistake is confusing teleoperated performance with autonomous performance.

1X makes this unusually visible. Its home robot NEO includes a scheduled “Expert Mode” in which a human expert can remotely supervise complex tasks the robot does not yet know. That can be a sensible bridge product: the customer gets more capability while the system gathers experience. But it also turns autonomy into an economic and privacy question. How often is the expert needed? What can the expert see? How quickly does the intervention rate fall as the robot learns?

The same logic applies to industrial pilots. A robot that completes 99 percent of a cycle but requires a person every few minutes may still be useful in a research program, yet expensive in production. Conversely, a slower robot that operates for an entire shift without rescue may create far more value.

For readers, this leads to a practical rule: never ask only “Can the robot do it?” Ask “How often can it do it without a human?”

10. Five questions to ask after every impressive humanoid video

Was the task autonomous? Look for teleoperation, scripted sequences, off-camera supervision or remote recovery.

Was the environment new? A policy trained in the same room may be excellent control but weak evidence of generalization.

How many failures were omitted? One successful take says little about success rate or intervention frequency.

How long did useful work continue? Minutes prove a skill. Hundreds or thousands of useful hours begin to prove a product.

What did the task cost? Robot price, maintenance, charging, supervision, integration and downtime determine whether automation is actually economical.

This checklist is deliberately boring. That is the point. The humanoid industry will mature when boring operational numbers become more informative than dramatic videos.

11. Safety becomes harder when the robot is allowed to improvise

Traditional industrial robots are often safe because we restrict them. They operate behind barriers, inside defined cells, on known paths. A general-purpose humanoid is valuable precisely because it is allowed to move through the same space as people and respond to situations that were not explicitly scripted.

That creates a tension between flexibility and predictability. A learned system may choose a different motion because a chair moved or a person stepped into the path. The behavior can be correct and still be difficult to certify with methods built for deterministic machines.

Industrial collaborative-robot safety is currently anchored by standards such as ISO/TS 15066. For service robots, ISO is in the process of updating ISO 13482, which covers safety requirements for robots used in personal and professional settings. The fact that the service-robot standard is being revised now is itself a sign that the deployment environment is changing.

A home robot also adds cybersecurity. It may contain cameras, microphones, maps of the home, account credentials and motors strong enough to move objects. A compromised smart speaker leaks information; a compromised mobile manipulator can change the physical environment. Security, permissions and local fail-safe behavior therefore become part of mechanical safety, not a separate IT concern.

The safest future robot may not be the one that “trusts its AI” most. It may be the one whose AI operates inside carefully designed limits, with independent systems able to stop motion, constrain forces and refuse actions when uncertainty is too high.

12. Do we even want androids that look human?

Science fiction tends to assume that once robots become capable, they will also become more human-looking. The research does not make that inevitable.

A 2026 systematic review of the uncanny valley and trust found that the relationship between human likeness and trust is highly context-dependent and that much of the evidence still comes from images, videos or hypothetical scenarios rather than long-term interaction with real robots. Another 2026 experiment found that anthropomorphism could increase emotional distrust in an industrial context.

That suggests a counterintuitive design path. A machine that works beside us may benefit from being understandable rather than human. Clear gaze direction, visible status lights, predictable movement and a voice that communicates uncertainty may be more valuable than synthetic skin.

There are settings — elder care, reception, education or companionship — where social presence may matter. But even there, looking human and behaving safely are different engineering goals. The first generation of truly useful robot assistants may therefore look less like androids and more like machines whose design has been optimized for cooperation.

13. The home is not simply a smaller factory

A factory is difficult, but it can be standardized. A home resists standardization. Every apartment has different furniture, clutter, lighting, objects, habits, children, pets and exceptions. People also care much more about mistakes. A scratched fixture, broken heirloom or discarded document may outweigh weeks of successful chores.

The home therefore requires a different kind of intelligence: not only task execution, but judgment about ambiguity. Should this cardboard box be recycled, or is someone keeping it? Is the glass on the table dirty, or is the owner still drinking from it? Is a closed bedroom door an obstacle to open, or a boundary to respect?

A general assistant needs a concept that industrial robotics can often avoid: permission. It must know not only what it can do, but what it is allowed to do. That means household preferences, identity, private areas, object ownership and uncertainty should become part of planning.

This is one reason the first genuinely useful domestic robots may be intentionally conservative. A robot that asks one unnecessary question is mildly annoying. A robot that confidently throws away the wrong object is unacceptable.

A household robot will not succeed because it can perform one impressive demonstration. It must work reliably among clutter, fragile objects, children, pets and constantly changing conditions — handling ordinary chores safely thousands of times.

14. Economics may decide the race before intelligence does

A humanoid does not need human-level intelligence to become economically important. It needs to cross a narrower threshold: the cost of useful automated work must fall below the cost or scarcity of the alternative.

That calculation is more complicated than the sticker price. A company has to count financing, integration, maintenance, batteries, floor modifications, software, supervision, downtime and the cost of failures. If a robot is cheap but needs frequent technicians, it is expensive. If it costs more upfront but works across three shifts and several tasks, the economics can reverse.

This is why Robotics-as-a-Service is likely to matter. Customers pay for capability or hours instead of buying an immature machine and assuming the technology risk themselves. The International Federation of Robotics notes that RaaS models are already common in professional service robotics. Humanoids fit that model particularly well because software and hardware are changing quickly.

The important economic metric may eventually be cost per autonomous useful hour. That single number forces the industry to confront everything marketing can hide: failures, human rescue, battery time, maintenance and actual productivity.

15. And what about jobs?

Humanoids matter to the labor debate for a different reason than conventional factory automation. A conveyor or welding cell automates a defined process. A general mobile manipulator is valuable precisely because the same hardware might be retrained to perform several physical jobs.

In the near term, the technology is much more likely to automate tasks than entire occupations. The first targets are repetitive material handling, parts movement, hazardous environments and work that is difficult to staff consistently. Humans remain necessary for supervision, maintenance, exception handling, quality decisions and the long tail of tasks robots do not yet handle.

Over a longer horizon, however, improved transfer could make physical automation more general. That is the point at which embodied AI starts to resemble the transformation already happening in digital work: the marginal cost of teaching the next task begins to fall.

That broader transition is explored in Next Horizon’s Life and Work in a World Where AI Does 90% of the Tasks. The key uncertainty for humanoids is not whether some manual tasks can be automated — they already can — but whether one physical platform can eventually absorb enough different tasks that automation stops being tied to a single machine.

16. Physical AI is the next step after AI agents

There is a useful continuity between the current AI boom and robotics. Generative models learned to create information. AI agents learned to use digital tools and pursue multi-step goals in software. Embodied AI adds sensors and actuators, allowing the same basic loop — perceive, plan, act, observe, correct — to operate in the physical world.

But physical agency changes the cost of error. A software agent can often undo a bad click. A robot cannot unbreak a glass, unhit a person or instantly restore an object it misplaced. This makes uncertainty estimation, permissions and verification much more important than they are in a chat interface.

For the software side of this transition, see Next Horizon’s Artificial Intelligence Explained: From Neural Networks to AI Agents. A useful analogy also comes from autonomous vehicles: self-driving systems taught the industry that a system can look excellent for 99 percent of the time while the remaining edge cases determine whether it is ready for the public.

17. What happens next: watch the milestones, not the promises

Predicting a year when “robots arrive” is almost meaningless because different forms of arrival are already happening. A better approach is to watch a set of measurable milestones over the next several years.

Longer autonomous intervals. The meaningful improvement is not another new task but fewer human interventions during ordinary work.

Cross-site transfer. A robot trained in one factory or home should become useful in another with little or no new data collection.

Dexterous reliability. Hands need to handle flexible, fragile and irregular objects with failure rates acceptable outside the lab.

Economically boring deployments. Robots should be purchased or leased because their cost per useful task works, not because the site is an R&D showcase.

Safety certification for adaptive systems. Standards and validation methods need to handle robots whose motion can change with context.

A shrinking teleoperation ratio. Remote rescue may remain, but the human minutes required per robot-hour should fall sharply.

Hardware price compression. Platforms such as Unitree already show that humanoid hardware can move down the cost curve; the question is how much reliability survives the compression.

Fleet learning. Improvements learned by one robot should become available to many robots without relearning every skill from scratch.

If these numbers improve, the industry can scale even if robots still look obviously mechanical and move more slowly than people. If they do not, better language models and more dramatic demos will not be enough.

18. So when do we get the science-fiction android?

Parts of it are already here. A machine can walk through a human workspace, understand spoken instructions, recognize objects, plan a multi-step task and manipulate the world. Research systems can transfer some behaviors across environments. Commercial systems are accumulating real operating hours.

What is missing is the combination. Fictional androids are not impressive because they can perform one difficult action. They are impressive because almost everything is ordinary to them. They have dexterity, common sense, memory, social judgment, energy endurance, reliability and the ability to learn new physical tasks without a team of engineers.

That combination remains far beyond what current public evidence demonstrates. The gap is easy to underestimate because a two-minute video compresses success while reality expands failure. A household robot may encounter thousands of small variations in a week; the product is defined by what happens on the one variation its developers did not anticipate.

So the milestone to watch is not “human-level robot intelligence.” It is something much more concrete: how many autonomous useful hours can the machine deliver, in environments it did not train in, before a person has to rescue it?

When that number becomes boringly large, the robot will stop feeling like a technology demonstration. It will become infrastructure.

The most important milestone for humanoid robotics may come when the technology stops attracting attention. A robot that quietly carries laundry, organizes objects or helps around the home without constant supervision would represent a deeper breakthrough than another spectacular demonstration.

Conclusion: the future robot will be judged by how little attention it needs

The humanoid field has finally moved past its oldest bottleneck. Building a machine that can stand, walk and recover from a push is no longer the entire research program. The center of gravity has shifted toward physical intelligence: perception, language, whole-body control, touch, transfer, error recovery and long-duration autonomy.

That does not mean a universal robot is around the corner. The strongest evidence in 2026 points in a more interesting direction. Humanoids are becoming useful first where the environment is partly structured and the economics can be measured. At the same time, foundation models are attacking the problem that prevented earlier robots from leaving those structured environments: every new task used to require too much new engineering.

The next decisive breakthrough may therefore look unimpressive on video. A robot completes the same shift every day. It needs fewer rescues each month. It learns a second task without a new controller. It handles a slightly different room without retraining. It recognizes uncertainty and asks before doing something irreversible.

That is the point at which “androids and robot assistants” stop being one more futuristic category and become a new layer of the economy — machines that can take AI out of the screen and give it a limited, useful, accountable presence in the physical world.

Next Horizon takeaway: The key question is no longer whether a humanoid can perform a task. It is whether it can perform useful tasks repeatedly, in unfamiliar environments, with a low intervention rate and a cost that makes sense. Reliability — not resemblance to a human — is the road from robot demo to robot assistant.

FAQ

Are humanoid robots already working in real factories?

Yes, but mostly in pilots, limited deployments and training programs rather than as a mass workforce. Figure, Agility Robotics, Boston Dynamics/Hyundai and others have reported real industrial operation or validation. The scale is still small relative to conventional industrial robotics.

Can you buy a humanoid robot for the home now?

Home-oriented products and reservation programs exist, including 1X NEO, but current systems should not be confused with fully autonomous household servants. Early products may depend on remote assistance for unfamiliar or complex chores.

Why make a robot humanoid instead of using wheels?

A humanoid is attractive when the environment and tools are already designed for people and the task mix changes. For flat, repetitive or tightly defined work, wheels, fixed arms and specialized machines are often cheaper and more efficient.

What is embodied AI?

Embodied AI is AI connected to a physical agent that perceives and acts in the real world. In humanoids it links cameras, touch and body sensors to reasoning, planning and motor control.

What is the biggest technical obstacle?

There is no single one. Dexterous manipulation, transfer to unfamiliar environments, long-duration reliability, energy, safety and the cost of collecting physical training data are all major constraints.

Will household robots need human faces?

Not necessarily. Humanlike appearance can sometimes help social interaction, but research on trust and the uncanny valley is mixed. Predictable behavior, clear communication and reliability may matter more than realistic skin or facial features.