Humanoid Robots: Why Useful Robots Are Still Years Away

The physical world creates an open-ended problem

Try a simple request: make coffee. The robot has to find the right machine, figure out how it works, pick an appropriate cup, handle grounds and water, and adjust if something goes sideways. Kitchens aren't standardized. Appliances vary. Cups break. Lids stick. Counters get cluttered.

The number of possible combinations grows fast. Train on one coffee maker and one set of cups, and the robot may fail the moment it sees something different. Train on many examples, and it gets better - but it's not realistic to collect enough data to cover every meaningful variation a real kitchen throws at you.

Robotics researchers have long responded to this variability by building systems around specific environments and specific tasks. That works in places like warehouses or factories, where setups are controlled. It gets expensive and fragile when the robot has to handle every home, every object, and every surprise.

Demonstrations are not the same as dependable performance

Robot videos can be helpful, but they often don't show how the demo was produced. The robot might be teleoperated, follow a carefully staged routine, or run in a space arranged for that specific test. Watching the end result doesn't tell you whether the task is autonomous, repeatable, and safe.

Even if the robot is acting on its own, a "looks impressive" success rate in a demo can be too weak for real operations. Boston Dynamics founder Marc Raibert put it bluntly: when a robot succeeds 70% of the time, for many jobs it's effectively not working. A business can't build a reliable workflow around a machine that fails nearly one out of three times. A homeowner probably won't risk hot pans or fragile dishes at that rate either.

None of this means demonstrations are useless. They show what systems can do under defined conditions. The mistake is treating a good trial as proof that a product can handle the full range of real everyday conditions.

Vision-language models help robots interpret scenes

Vision-language models (VLMs) process images alongside language. They can connect what the robot sees with a request - like identifying a cloth near a coffee spill or locating a laptop on a desk. This kind of contextual interpretation is useful because the robot doesn't need a unique hard-coded rule for every object it might face.

But recognizing an object doesn't automatically tell the robot how to act safely. Knowing "that's a cup" is different from choosing a grip that won't slip, lifting without spilling, and placing it where it belongs.

Vision-language-action models add movement

Vision-language-action models (VLAs) bring movement into the picture. They're trained on examples that link what the robot sees and what it's asked to do with the movements used to carry out the task. One way to collect those demonstrations is teleoperation: an operator controls the robot remotely while the system records what happens.

With that training, a VLA can attempt tasks like closing a laptop or wrapping up a headphone cable. Research systems (including Google DeepMind's Gemini Robotics) have also demonstrated more involved activities, like handling kitchen tools or assembling a simple lunch. Those are meaningful steps beyond "repeat one pre-programmed motion."

The catch is that a VLA's capability is still tightly linked to its training. If it hasn't encountered a relevant task, object, or variation, it may not know what to do next. As robotics professor Edward Johns has noted, doing several tasks doesn't automatically mean the robot can handle the space between them like a generalist would.

Why robot training data is hard to scale

Language models have benefited from a huge supply of existing text. Robots don't have an equivalent library of high-quality demonstrations covering the physical world. Researchers can collect demonstrations through teleoperation, learn from videos of people, or gather data from robots operating in real environments. Each approach has tradeoffs:

• Teleoperation produces directly relevant robot data, but collecting it takes time, equipment, and human operators.
• Human videos are everywhere, but they often don't include the precise movement, force, and sensor details a robot needs to reproduce an action.
• Real-world robot use can generate useful experience, but you can't easily deploy machines that aren't yet reliable or safe.

These sources can complement each other. Still, none of them removes the core problem: data collected for a finite set of tasks can't automatically cover every combination a robot will encounter outside the lab.

Simulation could make testing cheaper and safer

If the world model is accurate enough, researchers could test actions in simulation before trying them on real hardware. That could speed up development and reduce wear, risk, and the cost of collecting real-world data.

But simulation has to match reality closely enough to be useful. A simulated kitchen that gets friction, weight, soft materials, or object collisions wrong will teach the robot behavior that fails when it gets into an actual kitchen. Building a dependable connection between simulation and the real world is hard engineering on its own.

Early signs of compositional generalization

Some recent research hints at what broader capability might look like. Physical Intelligence's π0.7 system uses a lightweight world-model component that generates images of possible steps during a task. In one reported test, the robot was asked to put a sweet potato into an air fryer even though it hadn't been trained on that exact task. It made mistakes, but it managed a partial attempt.

That's interesting, not proof of a fully general robot. Researchers also found the system had seen small amounts of related data in training, including examples involving an air fryer basket. That could have helped it combine learned actions in a new way. People sometimes call this compositional generalization - using familiar skills in a combination that wasn't directly demonstrated.

The distinction matters. Combining a few learned moves is more flexible than following a fixed script. But it doesn't guarantee the robot understands the task the way a person does. The real question is whether this holds across different objects, environments, and instructions - and whether the robot can recover when something goes wrong.

Why factories come before homes

A factory or warehouse can be adjusted around a robot. Routes can be mapped, work areas bounded, and tasks standardized. And people can monitor performance and step in when something goes wrong. That makes it possible to measure whether automation is worth the operational cost.

Homes are less forgiving. They include different furniture, appliances, pets, children, clutter, and fragile objects. Putting away dishes sounds simple, but it requires recognizing where each item belongs, handling different shapes and materials, and adapting when cupboards are full. A domestic robot has to work in an environment that wasn't designed around its strengths.

Remote operation also complicates claims about home autonomy. Some products rely on a human operator to take over or guide the robot. That can be a reasonable way to develop a service or product. But it's not the same as a robot independently handling household work. In a private home, it also raises practical concerns - cameras, consent, data handling, and who can view the space.

Price is only one part of the business case

A robot priced at a few thousand dollars might sound affordable compared with a more expensive humanoid. But the purchase price doesn't tell you whether it's actually useful. You also have to factor in setup, maintenance, safety, supervision, and the value of the work it can complete reliably.

For businesses, the key question isn't whether a robot looks impressive. It's whether it lowers total operating costs or boosts output enough to justify buying and maintaining it. For consumers, a fair comparison isn't just the robot's price versus a person's time. It's the robot's price versus the value of tasks it can complete without frequent intervention.

That's why predictions of huge markets shouldn't be treated as proof of current demand. Hardware can get cheaper faster than capability becomes dependable. A low-cost robot that still needs constant oversight can end up being an expensive solution.