From factory floors to homes, the race is on to endow robots with the smarts of artificial intelligence.
They weld and paint cars parts, they lift heavy boxes and pallets, pick and pack items in warehouses, explore dangerous terrain, and deliver medicines by air. Robots or automated hardware systems that can grip and move, crawl, hop and even fly are by now firmly entrenched in the economy. Yet today's robots suffer from multiple constraints, mostly around the way they can see and understand the world around them and interact with it.
Until now. The rapid rise of AI systems and large language models (LLM) in particular is expanding the capabilities of robots and the expectations of their human handlers. If powerful and at the same time nimble machines get better brains and can learn faster from their environment, experts argue, it will unlock new opportunities to introduce robots to everyday settings in households, shop floors, care facilities or even hospitals. Humanoid bipeds or multi legged critters could soon work alongside humans or look after the elderly.
So far, the robotics boom has been constrained to industrial settings. According to the International Federation of Robotics, there are close to four million industrial robots in service worldwide, more than twice the number in 2017 and worth a total of USD 16.5 billion. China leads the way, followed by Japan, the U.S., Korea and Germany. Taken together, those five countries account for 80% of global installations, the IFR says in its latest report. The industry association also calls out two key trends to watch in 2025: generative AI, introduced to the world with ChatGPT, is quickly moving into the physical world by imbuing robots with broader understanding and learning abilities, plus humanoid form factors that turn robots into "general-purpose tools."
This prospect has attracted massive investments from established tech giants such as chipmaker Nvidia, Alphabet and Amazon to a slew of start-ups that focus on the goal to combine robotic brawn with adaptable brains such Agility or Figure AI. The narrower field of humanoid robots, long belittled as a money-losing quest, has become a hotly contested area. Companies from Tesla and ChatGPT parent OpenAI to Chinese carmaker BYD to consumer brands Meta and Apple are working on, or at least exploring, two-legged machines.
The hectic activity led one widely read AI market report to declare: "Robotics (finally) becomes fashionable (again) as the big labs pile in." Amazon, for instance, has already deployed more than 750,000 robots in its warehouses in the U.S. and uses generative AI for things like sequencing moves, traffic control, and how robots handle packages in unusual sizes.
Marco Hutter, director of the Robotics Systems Lab at the ETH Zurich and head of the RAI Institute's European office, considers the advent of LLMs as a pivotal moment for robotics. "It's a big step that will take some time to properly connect with the physical robotics world," he says. "LLMs and genAI have the potential to bring a higher level of intelligence and common sense to robotics, and achieve a natural human-robot interaction."
That interaction has, to date, between very circumscribed by the tasks that robots can perform. Every time a robot grips something or moves through a space crowded with other objects or humans, programmers would have had to think of the possible routines in advance and write appropriate code.
Teaching a machine how to make its way through a room or warehouse required prep work such as sticking guiding strips on the floor or running time-consuming mapping sessions - with often underwhelming results, as any owner of a robotic vacuum cleaner can attest to. "Our world is very challenging, highly unstructured, dynamic, diverse. Today, robots still lack in many ways when it comes to world understanding, autonomous mobility or dexterity skills," says Hutter.
The real advances are coming from software.
Language models that can understand plain English (or other languages) and quickly translate words into actionable code for robots have demonstrated there's another, easier way. "Robotics hardware is changing only very slowly. The real advances are coming from software," explains Jan Liphardt, associate professor for bioengineering at Stanford University. He compares AI to "a relative of yours who is blind and deaf. Imagine walking through a room, you're holding someone's hand and you're guiding them, describing the world and everything that's relevant. Once that information is transmitted and they have the right context, they can do amazing things."
Liphardt, who owns three robo-dogs he regularly takes on walks, argues LLMs work the same way. "Large language models can provide really compelling responses and commands to hardware systems. You only need a paragraph worth of information every few seconds to build a robot that has very interesting capabilities and behaves quite well," he says.
Plain language is but one part to capture one's surroundings. Robotic systems also rely on sensors such as cameras to see, hear and otherwise experience their environment. Observing humans (or other robots) in action lets them learn fast, as researchers from Johns Hopkins and Stanford recently demonstrated. They fed robots videos of surgical tasks, after which they were able to manipulate needles, tie knots and suture wounds, even correcting their own errors. The experiment could lay the foundation for robotic surgery that's no longer guided by a doctor controlling a joystick.
Technologists are offering an even steeper learning curve for robot brains in the form of entirely simulated environments. Such world maps are faster and cheaper to build and can be just as instructive, claim its proponents. Chip maker Nvidia, for instance, has introduced a 3D foundation model called Cosmos. Stanford professor Fei-Fei Li, often called "the godmother of AI," is also hard at work with a well-funded start-up called World Labs that aims to endow AI systems with "spatial intelligence as rich as our own."
Provide those training shortcuts to robots moving about on two legs or more, and the economic potential comes into view: increasingly smart software slipping into a body of its own and entering entirely new fields beyond factories and warehouses. That might explain why there are so many companies large and small investing in humanoid robots. By one count, there were more than 160 humanoid robot manufacturers worldwide in mid-2024. More than 60 of them are located in China, often tied to established electric vehicle manufacturers who can leverage their expertise with batteries and sensors, around 30 in the U.S. and another 40 in Europe.
There's also a downside. When robots become more capable, new security risks arise. As has happened with chatbots again and again, hackers can inject malicious prompts, but this time cause physical damage or harm humans. Networked swarms of machines could make such attacks much worse. "When one of them learns something, it can immediately transmit that skill or information to the other robots. A hacker attacking a swarm can create a lot of damage much more quickly," cautions Stanford's Liphardt.
Robotics development efforts ramping up around the world are part of the Fourth Industrial Revolution, according to Tobias Aellig, senior equity analyst with LGT Bank. "New technologies such as AI, the Internet of Things and advanced robots are driving productivity gains and operational efficiencies across nearly every sector of the economy," he says. Yet Aellig cautions there are still economic hurdles to overcome. "While robotics has seen significant advances in humanoid robots, mass market adoption remains a challenge as costs need to come down," he says. "In the meantime, human-robot collaboration continues to gain traction." Such "cobots" that assist humans with tasks such as heavy lifting or repetitive motions already account for 10 % of all new installations and are "growing much faster than traditional robots."
Care is one obvious application whose time has come, thinks Liphardt. "In my day job, I spend most of my time thinking about healthcare, and I wish patients would receive much better care." If robots were more independent and required less troubleshooting, they could assist an increasingly aged population. "It's already happening," says Liphardt and points to pairing chatty robots with seniors suffering from Alzheimer's oder Parkinson's disease. "If the choice is between your mother sitting in a chair and talking to herself all day, or having a piece of technology that puts a big smile on her face and gets her moving and interested - then an engaging robot is a good thing. Most people dramatically underestimate the extent to which humans can get deeply emotionally connected to a robot."
