The promise of physical AI

Created for physical artificial intelligence research, a pair of rotating human-like robotic arms mirrors the magician-like hand movements of a VR headset operator.

With enough training, such an arm could complete daily tasks on its own, says Tokyo company Enactic. The company is developing humanoid robots that wash dishes and do laundry in Japanese nursing care facilities that are understaffed.

Welcome to the future of AI, where AI begins to permeate the physical world in the form of smart robots, self-driving cars, and other autonomous machines.

Jensen Huang of US semiconductor giant Nvidia said last year that “the next wave of AI is physical AI.”

It’s “an AI that understands the laws of physics and can work among us,” Huang added, and an AI that understands “how we perceive the world.”

Tech companies are pouring billions into physical AI, and Morgan Stanley predicts there could be more than 1 billion humanoid robots in the world by 2050.

The buzz has only increased with videos of advanced androids, often made in China, dancing to Taylor Swift or effortlessly pulling heavy objects.

Beyond the promise of sci-fi robot butlers, this race raises concerns about unemployment, privacy, and how long it will take for these innovations to actually be useful.

Hiro Yamamoto is the 24-year-old CEO of Enactic. The company’s OpenArm physical AI training device is used by top universities such as Nvidia and Stanford.

He plans to begin deploying new robots currently in development next summer to “live with people in very chaotic, constantly changing environments” such as care homes.

“So it has to be safe,” Yamamoto said, with a soft exterior that won’t hurt anyone.

“Every Role of Human Beings”

In the Chinese city of Guangzhou last week, a female statue wearing a white fabric, like a fencer’s, with a glowing oval visor over her face walked slowly across a stage to cheers and whispers.

This was the latest humanoid robot released by XPeng, a Chinese electric car manufacturer that also focuses on physical AI.

Nimble machines made by American companies, such as Boston Dynamics’ robot dog, have been making headlines in recent years.

But government support and strong domestic supply chains are also helping Chinese rivals such as Unitree Robotics and EngineAI compete.

“I haven’t really thought about how many robots we’ll sell a year in 10 years, but I think it will be more than cars,” XPeng CEO He Xiaopeng told reporters.

XPeng’s robots can walk and even dance autonomously, but how well they handle objects, a more complex feat, has not been widely demonstrated.

Their dexterous fingers and flexible skin make them unlikely to be replaced by factory floor workers in China anytime soon, he said.

The cost of one robotic hand, which would need to be replaced regularly for heavy labor, could pay several years’ worth of salaries for Chinese workers.

But with enough data and training, AI humanoid robots could one day perform “almost any human role” from nanny to home cook to gardener, XPeng co-president Brian Gu told AFP.

On-the-job training

While text-based AI tools like ChatGPT are trained on vast amounts of words, physical AI models must also address vision and spatial relationships between objects.

For now, remotely controlling an AI robot and teaching it things like how to pick up a cup is “by far the most reliable way to collect data,” Yamamoto said.

Fine-tuning the “visual-verbal-behavioral” AI model requires just 30 to 50 demonstrations of each task, he added.

Enactic has approached dozens of nursing homes in Japan to let its remote-controlled robots take over menial tasks, freeing up qualified care workers to spend more time caring for the elderly.

Yamamoto said this field experience will train physical AI models that will allow robots to act autonomously in the future.

US-Norwegian startup 1X is taking a similar approach with its humanoid home helper NEO, which it plans to deliver to US homes starting next year.

The NEO costs $20,000 to purchase, but so far its performance has been spotty, with one video in the US media showing the robot struggling to close a dishwasher door, even when controlled remotely.

physical limits

In another embarrassing moment, Russia’s humanoid robot, said to be the country’s first, stumbled and fell flat on its face as it appeared on stage earlier this week.

Sara Adela Abad-Guaman, assistant professor of robotics at University College London, said there is currently a “huge gap” between robot AI systems and robot physical capabilities, which lags behind.

“Nature taught us that we need the right body to adapt to the environment,” Abad told AFP, citing the example of a mountain goat stumbling on ice.

Nevertheless, the big deals come even as a surge in investment in artificial intelligence fuels fears of a stock market bubble.

Japan’s SoftBank recently announced it would acquire industrial robot maker ABB Robotics for $5.4 billion, calling physical AI the “next frontier.”

Although automation raises questions about the future of human labor, Abad isn’t too worried.

After all, “our sense of touch is unparalleled,” she said.

© 2025 AFP