Humanoid robots can now pour drinks, jog, fold laundry, and even look convincingly useful in glossy demos. But put them in the real world, away from the carefully staged video and the friendly brochure copy, and the story gets messier fast: most still rely on teleoperation, narrow routes, or tightly controlled tasks, and that gap is now the industry’s biggest honesty problem.
That disconnect was easy to spot at Robotics Summit in Boston, where companies showed off robots that could greet visitors, carry coffee, or tidy a room. The reality, according to researchers and component makers at the event, is that the leap from impressive stunt to dependable helper is still a long one. A few years of bold demos have created the impression that general-purpose robots are basically here; the hardware and software say otherwise.
The demo is doing a lot of work
Elon Musk has been happy to showcase Tesla Optimus prototypes, including one that recently managed short, quick steps. Figure 03 has been presented as a tidy domestic assistant. Chinese systems from AgiBot and Matrix Robotics have been shown welcoming guests and serving coffee. But the more useful question is what happens when nobody is standing off-camera nudging the machine along.
According to Chris Matthieu of RealSense, many humanoids seen in public are either remote-controlled or locked into very specific tasks and paths. That makes the marketing sound more ambitious than the product. 1X Neo, promoted in October as a consumer-ready humanoid meant to change life at home, was also shown being operated by a human. The robot press has always loved a polished prototype; the bill usually arrives later.
AI is helping, especially with hands
The good news is that progress is real, and AI is the reason. William Okazaki of Renesas said the technology has accelerated the pace of improvement, and one of robotics’ long-standing pain points is finally getting less ugly: manipulation. Robots are now much better at grabbing objects with precision, and some sensors can even detect when they touch human skin.
Two model types are driving much of this shift. Vision-language-action systems combine written instructions with live camera input so a robot can connect what it sees with what it should do. ”World models” go a step further, training on huge amounts of images and video until they can predict what happens next in the physical world. That is a big step up from a machine that can only repeat a pre-scripted move.
- VLA models link instructions, camera vision, and action in one loop.
- World models try to predict how objects and scenes will change.
- Better grippers and tactile sensors are making physical interaction less clumsy.
Safety is the part everyone keeps underplaying
The industry’s real bottleneck is not just capability; it is trust. Charlie Kemp of Hello Robot said you only learn what a robot can really do when it tries something unexpected. Xinrui Bi of AgiBot pointed to the other familiar excuse: there still are not enough data. So companies are doing the obvious, slightly creepy thing and collecting more of it, placing cameras around homes, workshops, and factories to record human movement in the wild.
That would be a standard AI story if the machine lived in a laptop. It does not. A robot can bruise someone, drop something heavy, or misread a human body in close quarters, which is why Valentino Fagard of XELA Robotics stressed that any move into social spaces demands real safety around people nearby. The comparison to chatbots is flattering and useless; a bad answer in text is not the same as a bad decision with a motor behind it.
General-purpose robots are still years away
There are already robots working in the real world, but mostly in constrained trials. Boston Dynamics Atlas is operating at Hyundai, and Hexagon Robotics AEON is at BMW, though both are still in testing rather than being sold as normal commercial products. Daniel Fan of Innodisk said a robot that can handle general-purpose work will take more time, and that sounds less like a cautious disclaimer than a fair assessment of the state of play.
The bigger problem is that these systems are still black boxes. John Black of Brain Corp said models of this kind are nondeterministic, which means they do not behave exactly the same way every time. That is tolerable for a floor-cleaning bot in a supermarket aisle. It is a lot less charming for a humanoid expected to share a home, a factory line, or a sidewalk with people.
The next phase will probably be less about flashy walking clips and more about boring reliability: repeatable grasping, predictable motion, better tactile sensing, and real safety records. The first company to make a robot that is useful, affordable, and not mysterious will win a very patient market.