At ICRA 2026 in Vienna, humanoid robots were no longer a side conversation.
They appeared in booths, demos, competitions, research talks, startup pitches, and almost every discussion about what comes after industrial automation, collaborative robots, and mobile robots.
But the real question is not whether humanoid robots are impressive.
They are.
The more important question is:
Are humanoid robots becoming useful machines, or are we still watching advanced demos?
What We Saw At ICRA 2026
At ICRA, we recorded interviews with companies working on humanoid robots and dexterous systems, including AgiBot, Unitree Robotics, LimX Dynamics, and Xynova.
We also looked at other relevant participants, including PAL Robotics and Generative Bionics, to understand the broader direction of the field.
The result was not a simple story of “humanoids are ready” or “humanoids are hype.”
It was more interesting than that.
Humanoid robotics is moving quickly, but not evenly.
Hardware is advancing faster than deployment.
AI is improving faster than safety validation.
Demos are becoming stronger before applications are fully defined.
That is the state of humanoid robots at ICRA.
Not science fiction.
Not ready for every factory.
But real enough to take seriously.
Humanoid Hardware Is Becoming Available
For decades, humanoid robots were mostly research platforms.
They were expensive, fragile, difficult to maintain, and usually confined to labs.
That is changing.
At ICRA, one of the clearest signals was not a single robot doing a spectacular trick.
It was the number of companies treating humanoid hardware as a real product category.
AgiBot presented different robot families for different use cases: smaller robots for reception and interaction, wheeled platforms for factory environments, and bipedal humanoids for more general-purpose tasks.
That segmentation matters.
It is easy to say “humanoid robot” as if every humanoid is trying to solve the same problem.
They are not.
A wheeled humanoid in a flat factory is different from a bipedal humanoid walking through an unstructured home.
A research platform is different from a production machine.
A robot designed to carry boxes is different from one expected to interact with people in public.
Why The Market Is Still Early
Unitree showed another important part of the story: accessibility.
A few years ago, full-size humanoid robots were rare and difficult to access. Now, they are becoming platforms that research labs, integrators, and engineering teams can buy, test, and modify.
LimX Dynamics showed the same trend from another angle.
Its full-size humanoid platform is positioned for manufacturing, logistics, and service scenarios. However, the company also made clear that much of today’s demand is still research-driven.
That honesty matters.
The market wants to talk about deployment, but much of the industry is still building the tools, datasets, models, and control systems needed before reliable deployment can happen.
This is the first major shift:
Humanoid hardware is becoming available before humanoid applications are fully mature.
That is not a weakness.
It is how many technical markets develop.
First, the platform becomes accessible. Then researchers, developers, and early adopters discover what it can actually do repeatedly, safely, and economically.
The Hand May Matter More Than The Legs
Humanoid robots attract attention because they walk.
But useful work usually starts with the hands.
Walking across a booth is interesting.
Picking up an object, using a tool, handling a part, opening a door, connecting a cable, or loading a tray is much harder.
That is why Xynova’s focus on dexterous manipulation stood out.
The company presented its Flex 2 hand as a hybrid-driven dexterous hand using tendon drive and direct drive.
It also talked about degrees of freedom, backdrivability, compliance, CAN and EtherCAT communication, and the role of stable hardware in generating better data for AI training.
This is where humanoid robotics becomes more than a mechanical design problem.
If the hand is unreliable, AI cannot learn useful manipulation.
If the hand lacks sensing, the robot cannot understand contact.
If the hand is too rigid, it becomes dangerous around humans and fragile objects.
If it is too weak, it cannot perform useful work.
Dexterous Manipulation Is A Bottleneck
Dexterous manipulation is not just a feature.
It is one of the bottlenecks of the whole field.
At ICRA, the same theme appeared across research talks and workshops.
The conversation was not only about locomotion. It was also about embodied AI, manipulation, learning from demonstrations, world models, whole-body control, and contact-rich interaction.
That tells us something important.
The humanoid race is not only about who can build the most human-looking robot.
It is about who can build the most useful body for learning.
AI Needs Bodies, But Bodies Need Data
The last two years have changed the language of robotics.
Everyone now talks about embodied AI, physical AI, world models, robot foundation models, and vision-language-action systems.
Some of that language is useful.
Some of it is marketing.
The useful part is simple:
AI cannot remain trapped in text, images, and screens if we want machines to act in the physical world.
A robot has to perceive, decide, and move under real constraints.
It has mass, latency, actuator limits, and physical constraints.
It can slip, lose balance, hit objects, or break.
That makes robotics a very different problem from generating text.
The Data Problem Behind Humanoid Robots
At ICRA, the most serious conversations about AI were not abstract.
They were about data.
The key questions were practical:
- How do you collect enough real-world demonstrations?
- How do you transfer policies from simulation to hardware?
- How do you use teleoperation without creating an expensive data bottleneck?
- How do you train manipulation policies that generalize beyond a single object, table, or lab setup?
- How do you combine high-level reasoning with low-level control?
AgiBot’s presence was especially relevant here because the company is not only building robots.
It is also building around embodied AI datasets and development infrastructure.
The AgiBot World Challenge at ICRA was framed around the convergence of “Brain” and “Body,” with tracks for world models, VLM/VLA systems, and whole-body control.
That framing feels right.
The next stage of humanoid robotics will not be won by hardware alone, and AI will not be enough either.
A good humanoid robot needs the body, the data, and the learning system to evolve together.
Safety Is Still The Unresolved Question
The most important part of our interviews was not the hardware.
It was safety.
A humanoid robot is not a chatbot with legs.
It is a moving machine with mass, inertia, joints, batteries, arms, and sometimes a hand holding an object or a tool.
If it works near people, the safety problem becomes very concrete.
What happens if it falls?
What happens if a person steps in front of it?
What happens if it loses power?
What happens if perception fails?
What happens if a robot arm makes unexpected contact?
And what happens if the hand is holding something sharp, hot, heavy, or fragile?
Why Safety Changes Everything
In one interview, we asked a Unitree representative what would happen if a walking humanoid had someone in front of it and started to fall.
The answer included emergency stop, keeping distance from people, and obstacle avoidance through sensors such as LiDAR and cameras.
Those measures matter.
However, they also show the challenge.
An emergency stop is not the same thing as a complete safety case.
If a bipedal robot is dynamically balancing and you stop the actuators, the robot may still fall.
In human environments, the hazard is not only motion.
The hazard is loss of controlled motion.
AgiBot gave a more layered answer.
Its representative talked about sensing people in front of the robot, stopping on contact, force sensing, fall behavior, and soft materials designed to reduce harm.
The same interview included a striking reliability claim: according to the company representative, the time between falls had improved from roughly every 10 hours two years ago to around 10,000 hours based on fleet data.
That claim should be treated carefully.
It is not an independent certification.
It is not a published safety standard.
It is a company claim made in an interview.
But it points to the right metric.
From Demos To Real Deployment
For humanoid robots to leave the booth and enter factories, warehouses, hospitals, or homes, the conversation has to change.
It has to move from:
Look what it can do.
To:
How often does it fail, how does it fail, and what happens when it fails?
That is why safety standards matter.
Humanoids do not fit neatly into old categories.
They are not traditional industrial robots bolted to the floor, simple AMRs, or cobots.
Instead, they behave more like dynamically stable mobile manipulators operating in spaces designed for people.
That makes safety harder, not easier.
Europe Is Not Absent
There is a temptation to describe humanoid robotics as a China-versus-US race.
That is too simple.
China is clearly moving fast. Companies like Unitree, AgiBot, and others are pushing hardware availability, cost reduction, iteration speed, and manufacturing scale.
The United States remains central because of AI infrastructure, compute, foundation models, venture capital, and companies shaping the stack many robotics companies will use.
But Europe is not absent.
PAL Robotics, based in Barcelona, has been building humanoid and mobile manipulation platforms for years. At ICRA 2026, the company highlighted KANGAROO and TIAGo Pro, connecting dynamic bipedal robotics with mobile manipulation and research use cases.
Generative Bionics, from Italy, represents another European path: research transfer.
The company is connected to the long history of humanoid robotics at the Italian Institute of Technology and the iCub project.
At ICRA, its presence pointed to a broader question: how does public research become industrial capability?
That matters.
Europe may not always move fastest in hype cycles.
But it has strengths that are directly relevant to humanoids:
- Long-term robotics research
- Safety culture
- Industrial integration
- Mechatronics
- Public-private research ecosystems
- Experience with regulated environments
If humanoid robots are going to work in factories, logistics, healthcare, or public infrastructure, those strengths matter.
The Real State Of Humanoid Robots
After walking the booths, recording interviews, watching demos, and reviewing the broader ICRA context, this is our read:
Humanoid robotics is real.
But it is not one market yet.
It is several technical races happening at the same time.
One race is hardware: who can build reliable, affordable humanoid platforms at scale?
Another is manipulation: who can make hands and arms useful enough for real work?
There is also a race in embodied AI: who can collect enough data and train models that generalize outside the lab?
Safety is another race: who can prove these machines can operate near people without unacceptable risk?
And finally, there is the application race: who can find the first tasks where humanoids are not just impressive, but economically justified?
The industry is not waiting for all of these problems to be solved before moving forward.
That is why ICRA was so interesting.
The hardware is arriving, and the AI stack is starting to form.
At the same time, safety standards are catching up while applications are still being discovered.
Demos are getting better, but the questions are also getting harder.
That is usually a good sign.
A field is not mature when everyone agrees it is ready.
A field becomes serious when the questions stop being abstract.
At ICRA 2026, humanoid robotics felt serious.
Not because every robot was ready for deployment.
But because the conversation had changed.
The question is no longer only:
Can we build humanoid robots?
The question is becoming:
Can we make them useful, reliable, and safe enough to work in the real world?