Physical AI in business gets smarter and more autonomous
Leaping from the computer screen into dynamic real-world environments, physical AI and embodied robotics take on increasingly more complex and dangerous tasks.
Physical AI represents a significant leap forward from simple, repetitive automation tasks. Advances in foundation AI models enable robots and robotics software to perceive, reason, make decisions and operate autonomously in real-world business, healthcare and manufacturing environments.
Nearly 80% of organizations are engaging with physical AI, according to the 2026 survey of 1,678 senior executives by the Capgemini Research Institute, which also found that 60% of those surveyed believe physical AI will enable robotic applications previously impossible or impractical. The multinational bank Barclays estimates the humanoid robot market will grow from $2 billion today to $200 billion by 2035.
The overall physical AI market is even larger, projected to approach nearly $500 billion by 2030 as physical AI transitions from early pilots to scaled commercial deployment over the next three to five years, according to a report by consultancy Strategy&, part of PwC.
Physical AI, which includes embodied AI, is being deployed in numerous industries. In automotive, BMW used Figure AI's humanoid robots to help produce vehicles and work in the body shop. In shipbuilding, military shipbuilder Huntington Ingalls Industries uses physical AI for welding, assembly, surface prep, inspection and painting. In farming, John Deere plans to create a fully autonomous production cycle for corn and soybean farmers by 2030.
"We've had automation for decades," said Ben Wolff, president and CEO of embodied AI software maker Palladyne AI. "Humans have made all the decisions in advance when they build and program an industrial robot. But there are still millions of tasks that are dull, boring and sometimes dangerous that humans have to do because robots haven't been able to deal with unpredictable, unstructured or highly variable tasks and environments. That's where embodied AI comes in."
Wolff pointed to refurbishing aircraft parts. "Most often, old paint and corrosion are stripped off an aircraft part by a human wearing a hazmat suit operating a sand blaster or media blaster in a paint Robots are also being used when jobs vary and require flexibility. "Most manufacturers produce one part a million times, so it's very standardized, very consistent," said Marc Kermisch, chief technology and AI officer at digital manufacturing service provider Protolabs. "We produce a million [different] parts one time, so we have a lot of variability in how we work with our processes. We've trained robots and we're using AI to program those robots to be able to do basic tasks."
Moving to the edge
Edge AI enables robots to carry out tasks independent of network or cloud connectivity. "With physical AI, we're embedding that reasoning model … into the robot in and of itself," Kermisch said. "As a robot is using visual capabilities to inspect a part, it's able to make that reasoning right on the fly at the edge and be able to make decisions. This part meets quality; this part doesn't meet quality. That's all embedded, so it sits on the edge." AI training only needs to be within a specific area with localized knowledge, he added.
The [robots] that are working in the real world have gone more for a narrow than deep approach. They're focusing all the data efforts onto a single task.
Jack PearsonInvestment principal, RoboStrategy
"AI robots are independent," said Jack Pearson, investment principal at robotics investment firm RoboStrategy. "The big goal that everyone is trying to get to is general-purpose robotics, like one robot that can do lots of different things. The [robots] that are working in the real world have gone more for a narrow than deep approach. They're focusing all the data efforts onto a single task."
Physical AI can also increase productivity in smart factories, where they can "assemble, kit and ship finished goods with a high degree of autonomy," said Jeff Mahler, CTO and co-founder of robot system developer Ambi Robotics. "Physical AI primarily learns offline from pre-collected data on cloud training infrastructure. Self-learning is a key phase of the training process in which unlabeled data is used to learn a general representation of videos, images, sensor data and actions. This enables much better adaptation to new tasks, items and environments."
Sight + sound = perception
In addition to hardware for embodied robots, physical AI consists of many technologies. "You can't do much if you can't sense the environment with an AI system," said Angus Pacala, CEO and co-founder of sensor and perception systems provider Ouster. Optical sensors differentiates AI systems that can interact with the physical world from "chatbots and AI that's stuck in a computer console," he said.
"LiDAR and cameras are both optical sensors, but LiDAR is like the superhuman version that emits light and collects it," Pacala noted. "Most machines you see, whether it's a humanoid robot, a drone, a combine harvester, a forklift or a car, are using predominantly cameras and radar to navigate their environment safely and flexibly."
You can build the most sophisticated, human-like reasoning model, but if it doesn't have high-quality data feeding it, it's going to fail.
Dani CherkasskyCEO and co-founder, Kardome
While cameras help robots avoid obstacles, they don't help a robot understand who is talking to it or if it is even being addressed, said Dani Cherkassky, CEO and co-founder of voice AI provider Kardome. "Audio is the high-speed reflex layer that handles the critical reality of human intent before you ever need to trigger the heavy-lifting reasoning models," he said.
"Robots today are essentially context deaf," Cherkassky said. "We don't walk up to a human and stand three feet in front of them before asking a question. We don't expect machines to be so limited either." A machine that incorporates vision, LiDAR and audio understands its surroundings, he added.
It all comes back to the quality of data, however. "You can build the most sophisticated, human-like reasoning model, but if it doesn't have high-quality data feeding it, it's going to fail," Cherkassky said.
Physical AI is not suited for repetitive tasks, such as "jobs where if you're doing a million things, like a Coca-Cola bottling plant," Pearson said. "There's enough scale that you can just pay for a piece of fixed automation." He also pointed to a universal problem with AI systems, especially emerging technologies like physical AI and its various components. "It's a black box," he said, "so you don't know the process that happens inside. For critical safety tasks, the failure rates are still not perfect."
Chuck Martin, a New York Times bestselling author, futurist, speaker and columnist, has been a thought leader in emerging digital technologies for more than three decades.
