Humanoid Robots Advance, but General Purpose Autonomy Remains a Long Way Off

Humanoid robots are falling less often and can increasingly operate without a person following them around. That progress is visible at RoboCup, where 55 humanoid soccer teams competed and an 11-versus-11 demonstration game put the robots on the same pitch. But better locomotion and more convincing demonstrations haven’t solved the harder question: when does a human-shaped machine make practical sense, and what would it take for one to work reliably outside a carefully prepared setting?

Those questions were discussed by Joydeep Biswas of the University of Texas at Austin, Sanmay Das of Virginia Tech, Rina Dechter of the University of California, Irvine, Sabine Hauert of the University of Bristol, Marija Slavkovik of the University of Bergen, and AIhub’s Ioannis Erripis and Ella Scallan, in a conversation led by Lucy Smith. Their views converge on a distinction that matters to anyone evaluating a robot: a machine can show a new capability under specific conditions without being ready to take on a broad, unsupervised job.

RoboCup reflects a change in robot capability

Biswas described a sharp increase in humanoid capability over the past few years. At RoboCup, robots that once fell frequently began staying upright more often, and teams could deploy them without human handlers constantly nearby. He attributed the shift to several changes arriving together: improved hardware, better algorithms and new approaches to training.

One notable change is the use of learned controllers for walking. Earlier humanoid systems relied more heavily on classical control methods; newer systems can learn locomotion policies in simulation and then transfer those policies to physical robots. This sim-to-real approach helps teams train and refine behavior without relying exclusively on repeated physical trials. It is part of a broader change in how robots acquire movement skills, rather than a single component that explains the recent gains.

The competition’s scale gives a sense of how much activity has gathered around humanoid soccer. Biswas said 55 humanoid teams took part in the international competition, which also included a demonstration match with 11 robots on each side. RoboCup’s trustees traditionally play against the robots as a recurring measure of progress. Biswas said the comparison used to be easy for the human side; it is becoming less easy as the robots improve.

RoboCup is also changing the structure of its soccer competitions. Biswas said the leagues have been simplified and that, in 2028, the soccer competition will have only a humanoid league. The rationale is scientific: teams can study how perception, control and other functions come together in one integrated machine. A complete humanoid robot still isn’t perfect, but the organizers see enough progress to make the integrated form a productive subject for research.

That choice has roots in RoboCup’s earlier robot leagues. Hauert recalled becoming interested in robotics through the Aibo robot dog competition, working with Manuela Veloso’s team, before the field moved on to Nao robots and then humanoids. To her, the current emphasis is one more shift in the form factors that attract attention and help push robotics research into new territory.

She sees a similar change at major robotics conferences such as ICRA and IROS. The exhibition floors, she said, have shifted from a strong focus on drones to robot dogs and now to hands and humanoids, alongside industrial robots that have remained a fixture. The change is striking in the number of humanoid machines on display. Hauert is excited by the engineering work behind them, especially advances in control, but expects some of that work to benefit robots with other shapes too.

That transfer may be one of the field’s most useful outcomes. Better actuators, control methods, sensing and learning techniques can inform systems designed for particular jobs, even if those systems never look like people. For Hauert, humanoids belong in a larger robotics ecosystem; they need not become the default body for every application.

Task requirements should determine the shape

The appeal of a humanoid is easy to understand. Human environments contain stairs, handles, work surfaces and tools arranged around the human body. In principle, a robot with a similar shape could move through those spaces and use equipment made for people without requiring the environment to be rebuilt.

But that argument assumes a level of capability that present-day humanoids do not have. Biswas said there are few applications where a humanoid is clearly the best available robot today. For a simple pick-and-place job, for example, a wheeled robot can provide a stable base and a simpler planning problem. A humanoid has to manage its position while also controlling forces, velocity and acceleration across a more complex set of movements.

Every additional degree of freedom expands the choices a planner has to consider. That flexibility may be valuable when a task genuinely calls for human-like reach or balance, but it also adds control and planning demands. If the job is moving an object from one fixed place to another, a machine with fewer moving parts may be easier to build, operate and predict.

Slavkovik added a historical perspective on why human bodies are not a universal template for performance. She recalled examples in which people with different physical characteristics were considered better suited to particular tasks, including pilot Douglas Bader, who lost his legs in an accident. Her broader point was that a body’s suitability depends on the work. A robot’s shape should be judged in the same way: a human outline is one design choice among many, not proof that it is the best one.

Hauert imagines that capabilities associated with a person could eventually come from several specialized machines working together. A collection of robots with different bodies might perform the range of jobs a human can do, without forcing every capability into one complex machine. That arrangement would have its own coordination challenges, but it points to a useful design principle: build the body around the task, then consider how the system fits into the wider environment.

For some tasks, though, human shape may help with acceptance. Dechter suggested that people receiving healthcare assistance could feel more comfortable with a humanoid than with a box-like machine. The form may change how people interpret the interaction, even when it doesn’t improve the robot’s physical performance.

That advantage brings a risk. Das described a robot barista at a large technology convention that drew interest as it moved its arms to prepare coffee. He questioned the engineering trade-off: a semi-automatic espresso machine can make good coffee with a button press, so a complex robot is hard to justify on the result alone. Its human-like presence may be part of its appeal, but that appeal can lead people to attribute understanding and ability that the machine does not possess.

As humanoids become more capable, Das expects that tendency to matter more. People don’t respond to a humanoid in quite the same way they respond to a floor-cleaning robot or a robot dog. A human-like body invites human expectations. The team’s discussion raised the question of whether that makes interactions easier or creates an impression of competence and social understanding that the robot cannot support.

Hauert sees a related question in care settings. At the Bristol Robotics Laboratory’s assistive living research facility, researchers work with robots intended for human-robot interaction. A humanoid might fit some of those settings, she said, provided its design doesn’t mislead people about its abilities. In fact, a simpler, more cartoon-like machine may be safer and set more realistic expectations. She has seen videos of humanoids toppling over in crowds, a reminder that human-like appearance does not make a machine safe to approach in every situation.

Demonstrations don’t establish general capability

Videos of robots cleaning bathrooms, folding laundry or handling household objects can make broad autonomy seem close. The panel urged viewers to ask what the demonstration actually required: how many attempts produced the successful take, how much the setup was tuned to one scenario, and whether the robot could cope with a different tool, changed lighting or an unexpected object.

Biswas said many current systems are optimized for narrow conditions. A robot may succeed with a particular brush and a prepared room, but fail when the brush or surroundings change. The system may also depend on training methods that don’t scale well to the wide range of tasks people expect from a general-purpose machine. In his view, companies have financial reasons to emphasize the strongest demonstration, so engineers and researchers should be demanding about what a video does and does not establish.

Hauert added that a video can give viewers the impression that a task is already solved, even when the robot is remotely operated or finely tuned for the demonstration. She said there is no humanoid that can simply be unpacked and put to useful work in a place it has never encountered. The required training data and environmental preparation aren’t there. Biswas extended the caution to wheeled robots: that kind of out-of-the-box performance isn’t available there either.

Household robots advertised as broadly capable illustrate the gap. When Erripis asked about companies promoting home humanoids, Hauert replied that some are remotely controlled. Biswas said the machines aren’t able to perform all household jobs autonomously and that remote operation can serve as a way to collect training data. Hauert described a San Francisco company paying people to clean their homes while collecting data from the work.

That data collection may support development, but it shouldn’t be mistaken for autonomous performance. If a person is steering the robot through a task, the demonstration shows what the machine can do under human direction. It doesn’t show that the robot can perceive a new situation, choose a suitable action and recover from an error on its own.

Biswas compared enthusiasm for collecting robot demonstrations with an earlier idea in self-driving cars. Around 2014, some companies argued that building high-definition maps of the environment would be enough to solve driving. Detailed maps proved useful, but they didn’t resolve the whole problem. He sees a similar risk in assuming that a sufficiently large body of teleoperation data will produce general-purpose robots by itself.

The difficulty is the range of tasks and situations that a general robot would have to handle. The number of combinations grows rapidly, making it hard to gather demonstrations for every relevant variation. Biswas isn’t arguing that teleoperation data has no value. He questions whether scaling that data alone can cover the full range of circumstances a robot might encounter.

Slavkovik agreed that copying human actions has limits because people and machines solve physical problems differently. Birds and aircraft both fly, but their methods aren’t the same. Similarly, a robot doesn’t have to reproduce a human movement exactly to achieve a task; its own body may enable a different strategy.

Human video may still help. Hauert noted that there is a large amount of video showing people performing everyday activities. A humanoid body could make some of that material more relevant to robot learning, since the demonstrated movements are carried out by a body with a more comparable arrangement. In tests with vision-language models, she found that human-centered training data was harder to apply to robots that lacked a humanoid form.

Body shape is only one mismatch. Biswas pointed out that human videos are captured with sensors different from those on a robot. Even when the robot’s body resembles a person’s, it may not perceive the scene in the same way. The video data can be useful, but transferring an action from recorded human behavior to a physical machine remains a technical problem.

Learning requires interaction, not imitation alone

One limitation of demonstrations becomes clear when a robot makes a small mistake. An expert’s example shows a successful path through a task, but an imperfect machine may drift into a state that never appeared in the demonstration. It then needs to decide what to do next without a recorded example to copy.

Biswas argued that robots need experience from their own interactions. Watching a demonstration can provide an initial understanding, but the robot also has to try actions and learn from the results. That experience helps it respond when a task diverges from the ideal sequence. The point applies both to a physical robot and to the learning methods used to train it.

Dechter framed the distinction through causality. Observational data records what happened; interventional data comes from changing something and seeing what follows. To learn how an object or environment responds, a system needs to interact with it. Observation alone doesn’t reveal every cause-and-effect relationship relevant to action.

This creates a hard data problem. Language models can draw on huge stores of text, but robotics lacks an equivalent record of physical interaction. As Biswas put it, there is no “internet of physical interactions.” Robots have to learn in environments where actions have consequences, and the data gathered from those actions is costly and difficult to generalize.

There is also a mismatch between advances in language systems and physical skill. Dechter said it was striking to see robots begin speaking with large language models, as systems moved beyond generating text to voicing it. That can make a machine feel more human and more persuasive. But fluent speech doesn’t demonstrate that it can manipulate objects or navigate a room reliably.

Biswas pointed to Moravec’s paradox: tasks that seem straightforward to people, such as picking up a laptop, can be exceptionally difficult for a robot, while some forms of high-level planning may be easier to automate. People tend to judge difficulty from their own experience, but the demands on a machine are different. Language and reasoning progress can therefore create an inflated sense of how close physical autonomy is.

Hauert connected that gap to the difference in available data. There is vast material for training language systems; there is far less data that captures how objects move, how contact feels, or how actions succeed and fail in the physical world. The new abilities of chatbots don’t automatically transfer to a robot body.

The panel also raised the design choices behind the robots themselves. Erripis asked why many humanoids look menacing when consumer use is often part of their stated ambition. He contrasted them with Honda’s ASIMO, which had an approachable, toy-like appearance. Biswas agreed that the concern is valid and cited Fauna Robotics as a company working on more approachable designs.

He suggested that some teams are concentrating first on fine motor skills and the mechanical form needed to achieve them, leaving interaction design for later. Yet those choices are connected. A robot’s appearance shapes expectations, while its capabilities and behavior shape whether that appearance feels appropriate. Designing for interaction means considering both the machine’s physical performance and how people will interpret it.

Useful robots may arrive gradually

Hauert worries that attention and funding have shifted toward humanoids and “physical AI” at the expense of simpler robots that are already useful. A more complex body can be harder to verify and test, while a well-designed, less complex robot may already solve a bounded task. She wants the field to get working systems into applications instead of continually raising expectations for what the next form factor must accomplish.

Biswas outlined conditions that make mobile robots more manageable: a sufficiently detailed map, accurate prior knowledge of action costs, a clearly specified objective, a curated set of relevant entities, few surprises, and enough onboard sensing and compute. Manufacturing and logistics can often be organized around these assumptions. Remove one or more of them and the task starts to look more like open research.

Those conditions explain why a robot can perform well in a structured workplace yet struggle in a home. A factory or logistics site can define routes, objects and objectives; a household changes from one room and one task to the next. The challenge is not just walking or grasping. It is coping with uncertainty while meeting the reliability people expect from an autonomous product.

Hauert noted that the label “humanoid” is also being applied to machines with different configurations, including wheeled bodies and machines with six arms. As the category expands, it may come to encompass a range of forms, returning the discussion to the need for different designs. The label may attract interest, but it doesn’t tell a buyer whether a machine suits a particular job.

Slavkovik offered a qualified argument for the current excitement. A computer that could play chess didn’t solve general intelligence, but it drew attention and resources that contributed to later work in machine learning. Humanoids may likewise act as a catalyst for research that benefits other systems. She also warned that demonstrations can make capabilities look more complete than they are, especially when a compelling video is easier to sell than a careful account of limitations.

For Hauert, the meaningful breakthrough is not a particular body shape. It is seeing robots that work in everyday environments and watching research capabilities reach real applications. Biswas expects that progress to arrive through gradual adoption in limited, specific roles rather than a single sudden leap to a robot that can do everything. He noted that ChatGPT’s public-facing service helped bring existing language-model capabilities into public view, but he doesn’t expect robotics to have an equivalent moment that instantly reveals a fully capable machine.

Slavkovik’s personal test is laundry: a robot that can take unspecified laundry from a dryer and fold it into a neat stack, with up to a week to finish. Biswas responded that the task’s real difficulty lies in the conditions hidden inside the request. What kind of clothing? How accurately must it be folded? What happens when an item catches or the robot makes a mistake?

Those questions expose the difference between a successful research trial and a useful product. A robot that succeeds four times out of five might look impressive in a paper, Biswas said, but that performance would be unacceptable for a household product. Even a 99% success rate may leave too many failures if each one requires a person to intervene and rescue the machine. Consumers expect autonomy without having to spell out every exception.

There is no single date when humanoids will “work” in the way people imagine, because the phrase hides different tasks, environments and reliability requirements. RoboCup offers a specific long-term challenge: Biswas said its goal is for robots to defeat the human world champions by 2050. That is an open conjecture, not a schedule for everyday household autonomy.

The practical takeaway is to judge each robot against the job, the conditions of its demonstration and the amount of human help it needs. Humanoids are pushing hardware and control forward, and some of those advances will matter beyond humanoid machines. Their wider value will depend on whether researchers can turn narrow successes into systems that handle variation, recover from mistakes and earn the trust their human-like form invites.