Humanoid Robots: Better Mobility, Hard Limits on General Purpose Work
- AIhub
- 552 reads
RoboCup’s progress reflects advances in hardware and learned control, but household use still faces data, reliability and task-fit challenges.


Humanoid robots are becoming more capable, but a human shaped machine still isn’t a practical answer to most jobs. Recent RoboCup results show how better hardware, learned walking controllers and simulation training have improved mobility. They also expose the gap between a robot that performs in a prepared demonstration and one that works reliably in an unfamiliar home, factory or care setting.
That gap runs through the central questions facing robotics: which tasks benefit from a human form, what kinds of data can teach robots to perform them, and how should researchers and companies describe current capability? Researchers and practitioners 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 discussed those questions in a conversation moderated by Lucy Smith.
RoboCup reflects a change in robot capability
At RoboCup 2026, 55 humanoid soccer teams took part in the international competition. A demonstration game put 11 robots on each side. Biswas described a noticeable shift over the preceding couple of years: robots were falling less often, and teams could deploy them without people following every machine around to intervene.
He attributed the progress to several developments converging rather than a single breakthrough. Robot hardware has improved, algorithms have advanced, and training methods have changed. In particular, many contemporary humanoids use learned controllers for walking instead of relying solely on classical control. Simulation training followed by transfer to physical robots, often called sim-to-real, has become an important part of that development.
The competition has changed alongside the machines. RoboCup plans to simplify its soccer leagues, with only humanoid soccer remaining in 2028. The reasoning is that a single robot platform can now integrate functions such as perception and control well enough to make an integrated humanoid system a useful subject for scientific study, even though current machines remain imperfect.
RoboCup’s trustees have traditionally played against the robots as a recurring measure of progress. Biswas said the matches used to be trivial for the human players; they are becoming less easy as the years pass. The competition’s stated long-term challenge is for robots to defeat the human world champions by 2050. That is an open research goal, not a forecast that current systems will reach that level on schedule.
Hauert connected RoboCup’s development to a wider change in robotics. She began in the field through the competition’s Aibo robot dog league, working with Manuela Veloso’s team, and later watched the transition to Nao robots and humanoids. At major robotics events such as ICRA and IROS, she has seen exhibits shift from drones to robot dogs and, more recently, hands and humanoid platforms. Industrial robots remain part of the mix, but the number of humanoid machines on display has made the change hard to miss.
For Hauert, the significance reaches beyond humanoids themselves. Advances in actuation, control and manipulation can transfer to machines built for other shapes and tasks. A burst of work on humanlike robots may therefore benefit the broader robotics field, even if humanlike bodies do not prove to be the best design for most applications.
Form should follow the task
The argument for humanoids starts with the environments people have built. Homes, workplaces and tools are generally designed around human bodies. In principle, a capable robot with a similar reach and shape could use existing objects and move through spaces without requiring every environment to be redesigned.
The qualification is substantial: that argument assumes a functioning humanoid robot. Biswas said that, today, there are few applications where a humanoid is clearly the best solution. A robot tasked with picking up a block and putting it elsewhere, for example, may be better off with a wheeled base. Wheels provide a stable platform, and a simpler machine has fewer degrees of freedom to plan around.
A bipedal robot adds complexity. A planner must account for position, force, velocity and acceleration as it coordinates the body and manipulator. That enlarged set of variables makes a task harder to plan and execute than it would be for a stable, task-specific platform. A humanlike shape may offer flexibility in a human environment, but flexibility brings engineering costs.
Hauert argued for a diverse fleet of machines rather than one general form. Several robots with different bodies could work together to cover capabilities people associate with a human worker. In that model, generality comes from combining specialized machines, not from asking one humanoid to do every job. She still sees humanoids as part of the robotics ecosystem, and believes the attention they attract can push useful research forward.
Slavkovik offered a historical reminder that body shape affects performance. She cited pilot Douglas Bader, who lost his legs in an accident and was considered better able to handle high acceleration as a result. She also pointed to exploitative historical cases in which children were put to work because their smaller bodies suited certain tasks. These examples concern people rather than robots, but they underline a broader design point: a body’s suitability depends on the task, and a humanlike adult form is not a universal engineering optimum.
The same practical logic appears in everyday automation. Das recalled seeing a robot barista at a large technology convention. The machine moved its arms to prepare coffee, but in his view it did not make coffee as well as a semi-automatic espresso machine operated with a button. The spectacle of a robot performing a familiar human job can be compelling; it does not establish that the robot is the most effective way to do that job.
Human appearance changes how people respond
A humanoid’s shape may matter even when it offers no mechanical advantage. Dechter suggested that a humanlike robot could create a more comfortable impression in social applications such as healthcare assistance. People may respond differently to a humanoid than to a box shaped machine, and that response could be part of the use case.
But appearance also creates expectations and risks. Das warned that people are more likely to attribute human qualities to humanoid machines than to a Roomba or a robot dog. As robots gain conversational and physical abilities, users may treat them as more understanding or capable than they are. That can affect trust and how people rely on them.
Dechter said that large language models have made this issue more immediate. Systems that generate spoken language can feel more human than software that only returns text. A machine that looks and sounds human may be especially persuasive, even when its underlying understanding and physical competence remain limited.
Hauert’s work at the Bristol Robotics Laboratory’s assistive living research facility focuses on robots designed for human interaction. She sees room for humanoid forms in care settings, provided the machines do not mislead users about what they can do. The form can set a high bar: people may expect a humanoid to be more capable than a less humanlike device. Simpler, more cartoonlike robots may sometimes be a safer fit, especially in settings where a machine could topple among people.
That leaves designers with a joint problem. They need to consider motor capability and human response together. Building a body first and postponing questions of interaction can produce machines that move impressively but feel threatening or invite misplaced confidence. Conversely, a friendly appearance cannot compensate for unreliable assistance.
Science fiction may help explain why the humanoid shape attracts so much attention. Slavkovik suggested that stories have long trained audiences to imagine robots as humanlike characters. A box shaped device needs explanation before it can become the protagonist of a story; a face and limbs communicate a familiar role immediately. This cultural bias may shape what companies build and what audiences expect from robotics.
The recent rise of large language models adds to that expectation. If AI can converse fluently, a humanoid robot can look like the next step toward general intelligence. Hauert said people often press for a date when humanoids will work as expected, as though the direction were already settled. In practice, the definition of “useful” or “capable” matters more than the label.
Demonstrations don’t establish general capability
Promotional videos of robots cleaning bathrooms or folding laundry can make a task look solved. Biswas argued that the field needs greater truth in how it presents such demonstrations. A video rarely reveals how many attempts were required, how carefully the environment was prepared, or how narrowly the machine was tuned to the particular brush, lighting and object arrangement.
Those details matter because a robot can succeed under tightly controlled conditions and fail when a small variable changes. A different tool, altered lighting or an unfamiliar room can expose limits that a polished clip hides. Biswas said current systems are often highly optimized for specific scenarios, rather than being general purpose. The training methods behind them also do not yet scale easily to the broad range of tasks people expect.
Hauert noted that viewers may assume a capability is ready for use once they have seen it in a video. Some demonstrations involve remote control; others depend on extensive tuning for one setting. There is no general “out of the box” humanoid that can be put into an unfamiliar environment and expected to do useful work. The data needed to train for such settings is not yet available at the required scale.
Biswas added that the limitation applies even to wheeled robots. The obstacle is not uniquely caused by humanoid bodies. A mobile robot that works reliably in a carefully bounded space is still far from a machine that can handle the unexpected conditions of ordinary life.
Household robots advertised as humanoid assistants illustrate the distinction. Biswas said none can currently perform all household tasks autonomously; some are remotely controlled as part of data collection. Hauert mentioned a San Francisco company that pays people to do household cleaning while gathering data in homes. When a human guides or performs the work, the footage may help train a future system, but it does not show that the robot already performs the task independently.
There is a business reason for the gap between demonstration and deployment. Companies benefit from showing their platforms at their best, while failures are less visible. Biswas said the influx of companies and financial incentives can pull against scientific progress when a striking demo is valued more than evidence of reliable performance. For engineers evaluating a system, the useful questions are concrete: Was it autonomous? How many trials succeeded? What changed between trials? What conditions were fixed?
Why language AI does not transfer directly to robot skills
Public expectations can also blur different kinds of intelligence. A chatbot can plan or explain a task in language, but a robot must sense and act in a physical environment. Biswas invoked Moravec’s paradox: some high-level reasoning tasks that seem difficult to people can be comparatively straightforward for computers, while seemingly simple physical skills remain hard for machines.
Picking up a laptop may look trivial to a person, but the robot has to perceive the object and its surroundings, position its body and gripper, make contact without slipping or damaging the object, and adjust if the action does not go as planned. Language models draw on enormous stores of text. Robotics has no equivalent “internet of physical interactions” containing the varied sensory data and consequences needed to learn everyday manipulation.
Data collection through teleoperation is one attempt to close that gap. A person controls a robot through a task, creating demonstrations that can be used for training. Biswas acknowledged that these records are useful, and said companies are spending from millions to billions of dollars on teleoperation efforts. Researchers disagree, however, about whether scaling up demonstrations alone can produce general purpose robots.
Biswas doubts that it can. The number and variety of tasks people expect robots to handle grow combinatorially, making it unlikely that demonstrations could cover every relevant situation. He compared the current enthusiasm for collecting demonstrations with an earlier period in self-driving car development, around 2014 or slightly before, when some companies expected detailed high definition maps of the environment to be enough to solve driving. That approach proved insufficient on its own. Data collection can help without being a complete solution.
Slavkovik made a related point through the difference between biological and engineered movement. Birds and airplanes both fly, but they do so through different mechanisms. A robot does not necessarily need to reproduce the human movement that generated a demonstration. It may be more effective to learn an action suited to its own body.
Human video is another possible source of training material. People have recorded vast numbers of everyday activities, and a humanoid robot may have a physical frame that makes some of those observations easier to map onto its own actions. Hauert said tests with vision-language models found it harder to use human centered data with robots that do not have humanoid forms.
Even a close match in body shape would not remove the sensing gap. Biswas pointed out that human videos were captured with different sensors from those used by robots. The camera angle, available measurements and way the data was collected all differ. Human video may be valuable, but translating visual observations into reliable robotic action remains an open problem.
Robots need experience, not imitation alone
Several participants emphasized that robots need to learn by acting. A demonstration can provide an initial model of a task, but a robot that copies an expert action may make a small error and enter a state that never appeared in the demonstration set. It then has no recorded example of what to do next.
Interaction lets a learning system encounter those departures and build information about what actions produce which outcomes. Dechter described the distinction in causal terms: observational data shows what happened, while interventional data comes from acting and seeing what changes. Observation alone cannot reveal every cause and effect relationship needed to choose a response.
Biswas said that learning through interaction can also adapt behavior to a robot’s particular shape and capabilities. A human demonstration may show the goal, but a machine can develop a different movement strategy that better suits its limbs, sensors and control limits. The challenge is to make this learning effective and safe enough to use in the real world.
That requirement complicates the usual push for more data. Teleoperated examples are often collected by a human who already knows how to complete the job. Real autonomy requires a robot to recover from slips, unexpected objects, changes in lighting and other conditions that were not present in the demonstration. Better learning methods will need to handle those variations, not simply repeat successful trajectories.
Established robots still have a role
Hauert is concerned that attention to humanoids could shift the definition of progress toward more complex machines, even where simpler robots already work well. Earlier mobile robots had useful applications when equipped with sufficient onboard computation and good sensing. A humanoid body can make a system harder to test and verify, and less environmentally efficient, while prompting researchers to chase capabilities that a simpler machine may not need.
The pattern can affect which projects receive support. Hauert said humanoid and “physical AI” work is attracting attention and funding, while teams outside that wave may find it harder to raise money. Her concern is that the field could lose practical value if established forms and working systems receive less attention simply because they are less visually compelling.
Biswas described how much a successful mobile robot depends on assumptions about its operating environment. It needs a sufficiently detailed map, accurate predefined action costs, a clear task objective, a curated set of relevant objects, few surprises, and adequate onboard sensing and computation. Manufacturing and logistics sites can often provide those conditions. Remove them, and the machine moves into a much harder research problem.
This distinction is useful when judging claims about generality. A robot that succeeds in a mapped warehouse with known objects can still be valuable. Its usefulness depends on matching the system to the conditions it can handle, rather than treating performance in a bounded setting as proof of general competence.
Hauert also observed that the word “humanoid” is being applied to an expanding range of designs, including wheeled machines and robots with six arms. The category may eventually encompass many body plans. That could bring discussion back to the central engineering question: what configuration suits the job?
Slavkovik compared the role of humanoids in public attention to early chess computers. A machine that plays chess does not by itself solve general AI, but a compelling demonstration can draw resources toward broader research. She cautioned that dramatic depictions of robot capability can cross into deception, even if spectacle helps sell the field. Attention may produce useful work, but that outcome does not excuse unclear claims.
Expect gradual deployment and demanding reliability tests
There may be no single public launch that marks the moment robots suddenly become useful in the way ChatGPT’s public website brought language models to a wide audience. Biswas expects more gradual adoption: robots will take on limited capabilities in particular scenarios, with deployment expanding over time. That path is less dramatic than an all purpose robot, but it fits the way capability is being built.
For household work, laundry offers a particularly clear test. Slavkovik said a robot that could take unspecified laundry from a dryer, fold it into a neat stack and finish within a week would impress her, whatever its price. The deliberately loose deadline leaves room for slow operation; the difficult requirement is succeeding across different kinds of clothing without hidden conditions.
Biswas pointed to reliability as the hard part. A system that handles most attempts but fails badly every so often can leave its owner with a failed task and a robot that needs rescuing. Research papers may describe an 80% success rate as strong, but a consumer product with an 80% success rate would be a poor fit for a task people expect it to complete unattended. Even 99% may fall short when failures are costly or inconvenient.
That is why “useful” needs a defined operating envelope. A robot might fold a particular garment in a cleared space under good lighting, or it might manage varied laundry without supervision. Those are very different capabilities. The same applies to cleaning, coffee preparation and care work: performance claims become meaningful when the task, setup, level of autonomy and failure rate are clear.
Robot appearance deserves similar care. Erripis asked why many humanoids look menacing despite ambitions for consumer use, recalling Honda’s ASIMO as an approachable, almost toy-like machine. Biswas agreed that friendly designs exist, including Fauna Robotics, though he described that approach as an exception. He suggested that companies have focused first on solving fine motor control and have treated human interaction as a later concern. The two should be designed together if robots are meant to work closely with people.
The clearest sign of progress may not be a humanlike machine completing a staged task. Hauert said she would look for robots working in everyday environments and for research advances reaching real applications, regardless of the machines’ shapes. RoboCup demonstrates genuine movement in hardware and control, but broad autonomy remains a longer project. For now, the strongest engineering choice is still to match each robot’s body, training and reliability to the job it must perform.