AI is Supercharging the Race for General-Purpose Robots

In a world where self-driving robotaxis glide through major city streets without drivers behind the wheel and delivery drones autonomously fly through the skies to drop off orders at customers’ homes, the idea of general-purpose robots helping humans with various tasks in workplaces or even homes may not seem far-fetched.

But that future hinges on developing increasingly autonomous robots powered by modern artificial intelligence—an ambitious vision that has motivated many researchers to become startup founders while also attracting billions of dollars in investment.

“When I started maybe about 15 years ago, I led a project team that was focused on autonomy, but in that era, the goal of that team was to just get a robot to navigate from point A to point B,” said Matt Malchano, vice president of software at the robotics company Boston Dynamics based in Waltham, Massachusetts. “And now, when we think of autonomy, we think of this huge space of tasks and things that we can imagine a robot doing on its own.”

It was previously difficult to imagine a practical path for creating general-purpose, autonomous robots like housekeeper Rosie from The Jetsons or the various droids like C-3PO from Star Wars, especially when robotics labs and companies were still struggling to solve autonomous navigation and even self-balancing, in the case of walking robots. In 1979, the experimental autonomous vehicle known as the Stanford Cart required five hours to successfully move 20 meters through an obstacle-filled room. The first bipedal robot capable of walking on its own without losing its balance was developed in 1996.

But robot autonomy has always been a “moving target,” with the goal of reaching a point where robots can do an increasingly larger subset of things that humans can already do, ideally without direct human supervision, Malchano told Ars. The International Standards Organization defines autonomy in robotics as the “ability to perform intended tasks based on current state and sensing, without human intervention.”

More recent advances in AI—such as reinforcement learning in the 2010s and large foundation models trained on huge amounts of data in the 2020s—have “unlocked” the ability to “imagine a world where the robot can do sequences of activities and really understand the tasks, and that’s very exciting,” Malchano said. Now, multiple research labs and companies are racing to develop general-purpose robots capable of handling a wide variety of tasks independently in more complex, unpredictable environments.

Such robots will not necessarily be humanoid in appearance and function despite the substantial amounts of investor money going into humanoid robots. But whatever their form, they could represent a significant step beyond the millions of industrial robots and service robots that already perform specific tasks within the relatively controlled environments of factories and warehouses.

“There’s an assembly line, the robot is supposed to do a particular motion, and if you do that motion reliably and repeatedly, that’s a basic factory level of autonomy,” said Sergey Levine, a computer scientist at the University of California Berkeley and cofounder of the AI and robotics company Physical Intelligence. “The next level, though, the one that is currently at the edge of what’s possible—like a research topic that’s making its way into the real world—is that the robot can do a thing in an unstructured environment reliably.”

Ars interviewed robotic researchers and founders about how AI has supercharged interest in robotics, the challenges of making general-purpose robots, how safety is a make-or-break issue for robot workers, why surgical robots still have limited autonomy, and when to expect robot helpers in people’s homes.

The modern AI impact on robotics

Levine’s startup, called Physical Intelligence, is working toward achieving practical robotic intelligence that can empower many different types of robots operating autonomously in open-world environments. “I don’t think it will be the one ultimate robot, like a super advanced humanoid that can do everything,” Levine told Ars. “I think it will be a general AI model that can power lots of different robots that are well-suited for their job.”

For example, a small robotic arm hanging from the ceiling might prove more suitable for a tiny New York City apartment, whereas a “hulking giant robot” that moves heavy objects could be more handy on a farm, Levine suggested. “I’m sure we’ll also have good humanoids, but there will be other stuff, too,” he said. “There will be whatever makes the most sense for the job.”

But developing autonomous robots that can operate more independently in the open world involves “step changes in technological complexity,” Levine said. He described such robots as needing to handle complex environmental perception, requiring robust motor skills and the ability to overcome basic mistakes and process instructions from humans. In addition, such robots will need to learn how to generalize their behaviors to handle new situations.

Many researchers are trying to make this happen through a combination of AI techniques involving reinforcement learning and large pre-trained models, Levine said. Reinforcement learning involves training robots to perform specific tasks through trial-and-error, whether using physical robots interacting with the real world or computer simulations. At the same time, foundation models pre-trained on huge amounts of data—such as visual-language models trained on images and text—can provide some basic prior knowledge of the world to help robots react more appropriately and avoid unnecessary mistakes in various situations.

“Reinforcement learning is like how, after you practice your tennis swing many, many times, you can get really good at it,” Levine explained. “But to get there, you first need to have a kind of basic common sense to even get started.”

That combination of AI techniques and the gradual expansion of accessible training data—such as humans teleoperating robots to show how to perform specific tasks—has enabled “encouraging” progress in training robots to perform many different tasks reliably under various conditions, Levine said. “The key to making modern machine learning systems work is to get enough of a critical mass of data so that we see generalization,” he said.

But there is still a widely recognized data gap when it comes to collecting enough of the right data for training robots to perform physical tasks. The more costly and time-consuming methods involve humans wearing teleoperation rigs to directly guide a robot’s physical motions in training or running many experimental trials with robots in labs or other environments. Manually coded simulations grounded in physics can help train robots more cheaply in virtual environments but can fail to capture many real-world complexities and uncertainties.

Robotics researchers have also been developing world models to help robots predict the consequences of their actions in the physical world and plan accordingly. Some implementations of such AI models are trained on primarily visual data to learn how physical environments work, with some companies even collecting first-person videos for training data by hiring gig workers to wear head-mounted cameras as they do household chores or other tasks.

This approach to robot training is cheaper than running real-world experiments with robots, but world model development remains computationally expensive and may also struggle to accurately replicate real-world physical interactions.

For now, general-purpose robots remain on the horizon. Current training methods, such as reinforcement learning, can produce robots that are very good at consistently performing a specific task under very specific conditions—but such robots may falter on the same task under different conditions. Meanwhile, the growing variety of data from teleoperation and other methods can train robots to learn a “diversity of tasks, but not to the 99.99 percent level of robustness,” Levine said.

“You can either have something that is kind of OK but not amazing at everything or something that’s extremely good at one thing,” he said. “We really want something that’s extremely good at all things, and that’s still at the frontier of research.”

Learning through work

Fortunately, the world won’t need to wait for the development of general-purpose robotic capabilities to find practical uses for robots. Many companies have already been developing and selling specialized industrial and service robots for decades—and both newer startups and more established robotics companies are already applying the latest levels of robot autonomy toward handling a growing array of tasks.

One of the most well-known contenders is Boston Dynamics, which originally spun out of an MIT lab in 1992. The robotics company is popularly known for its viral video demonstrations of quadruped and biped robots, with the most recent examples including its Atlas humanoid robot learning various soccer moves during the 2026 World Cup.

But for several years, Boston Dynamics’ four-legged Spot robot has been conducting autonomous inspections of facilities that are more hazardous for humans, including Massachusetts converter stations operated by the electricity and gas utility company National Grid and culvert pipes running beneath California highways.

Such robotic autonomy is “really about the ability for that robot to navigate through an environment and perform actions on its own, taking photos and sensor measurements of that environment,” Malchano said. “That’s a capability that we’ve packaged into a product and we just sell to be used by people who are not roboticists but instead are focused on making their facilities really work great and not break down—that’s a form of autonomy that’s available today.”

One particular challenge for Spot involved learning to walk on slippery floors in customer facilities, which required additional training through reinforcement learning. “We retrained how that robot chooses to walk and its ability to recognize that it’s on a slippery floor and then take actions to stay stable and be able to navigate across that in a similar way to how a human might walk across ice,” Malchano explained.

At the same time, the company’s wheeled Stretch robots with large robotic arms have been handling large boxes and packages in warehouses operated by logistics companies such as DHL. “We’ve continually adapted it to different forms of packages, the ways that trucks are loaded, and the structures of the trucks themselves that we’ve encountered by interacting with the real world,” Malchano told Ars.

Boston Dynamics is also currently ramping up manufacturing of the all-electric Atlas humanoid robot. The humanoid robot is undergoing training and testing at the Robot Metaplant Application Center run by its South Korean parent company, Hyundai Motor Group. The goal is to have trained Atlas robots performing tasks at the Hyundai Motor Group Metaplant America—a huge electric vehicle factory located in Ellabell, Georgia—by 2028.

“I think we’re very lucky to be affiliated with Hyundai, which obviously produces fantastic cars that are produced at scale,” Malchano said. “Being able to leverage that capability to get to scale is really important for building robots.”

The Hyundai and Boston Dynamics effort aims to have the capacity to produce 30,000 humanoid robots annually by 2028. Whether the world can find a use for so many humanoid robots depends on how useful and cost-effective they can be in traditionally human workplaces. There are also initial signs of pushback from Hyundai’s own human workforce—the Hyundai Motor labor union approved a potential strike on June 25 as it negotiated with the South Korean automaker about job protections related to the upcoming Atlas robot deployment.

But for now, general-purpose robots that approach human levels of flexibility and adaptiveness remain many years away. “We’ve come to expect that if you ask a person to do a task, they’ll do it right with some training almost all the time,” Malchano said. “I think we’re still understanding what it takes to achieve that level of reliability for general-purpose, AI-driven tasks.”

The roadmap from warehouses to homes

Future progress in robotics probably won’t look like a “ChatGPT moment,” in which robots trained on enough data suddenly become capable, said Jonathan Hurst, cofounder of Agility Robotics and a robotics researcher at Oregon State University. He said collecting training data for robots isn’t as straightforward as scraping Internet text, images, and videos—it requires collecting much more real-world data related to controlling and coordinating all the robot’s joints and limbs as it interacts with the physical world.

“It’s dramatically harder to have an embodied AI; it’s 10 times harder to have an embodied AI,” Hurst told Ars. “The data that enabled large language models doesn’t exist and never will exist for embodied AI.”

Hurst anticipates robots making gradual progress over the next several decades that allows them to enter more generalized environments and situations. His company, Agility Robotics, was the first to earn a long-term commercial contract for humanoid robots by deploying its Digit robots at a GXO logistics warehouse in Atlanta, Georgia, starting in 2024. Inside the warehouse, Digit robots crouch to lift and move totes filled with products from order-picking areas to conveyors.

Agility has since deployed more humanoid robots to work on the automotive production lines of Toyota Motor Manufacturing Canada and the South Carolina factory of German automotive manufacturer Schaeffler. It has another commercial agreement to integrate Digit robots into a facility of the e-commerce company Mercado Libre in San Antonio, Texas. The robots have even done testing in Amazon warehouses.

On June 24, Agility announced plans to become the first “pure-play humanoid company” publicly listed on a major North American exchange. The company says its robots have already accumulated more than 65,000 hours of operations through the initial commercial deployments and pilot programs.

Digit’s workplace duties have mostly focused on picking up and handling totes and bins. But the next step could be picking up items and putting them into bins, then handling cardboard boxes of many different sizes and shapes, Hurst said. That could eventually lead to work opportunities in the back rooms of retail and grocery stores, which would be less structured environments and potentially more chaotic than factory assembly lines or warehouses.

Further down the road, humanoid robots such as Digit may start riding around in autonomous vehicles to deliver packages to people’s front doors. The very last stop might involve directly helping people in their homes. But Hurst cautioned that robotic development is still several