Physical AI — robots and autonomous machines that operate in the real world — is attracting heavy investment and fast development. Companies from Tokyo to Guangzhou are training humanoids for chores in care homes and homes, often using teleoperation and as few as 30–50 demonstrations per task to teach vision-and-action models. Experts warn that physical limits, expensive components and safety concerns mean real-world performance remains uneven, even as forecasts project up to a billion humanoid robots by 2050.
Rise of the Robots: Physical AI Moves from Lab to Living Room
Physical AI — robots and autonomous machines that operate in the real world — is attracting heavy investment and fast development. Companies from Tokyo to Guangzhou are training humanoids for chores in care homes and homes, often using teleoperation and as few as 30–50 demonstrations per task to teach vision-and-action models. Experts warn that physical limits, expensive components and safety concerns mean real-world performance remains uneven, even as forecasts project up to a billion humanoid robots by 2050.
08:34 AM, 11/14/2025Technology
Rise of the Robots: Physical AI Moves from Lab to Living Room
A pair of swivelling, human-like robotic arms built for physical artificial intelligence research mirror the motions of an operator wearing a VR headset, waving hands like a magician. Tokyo startup Enactic says that with enough teleoperated training these arms can learn routine tasks — the company is developing humanoid machines to wash dishes and do laundry for understaffed care homes in Japan.
What is physical AI?
Physical AI refers to systems that act in the material world: smart robots, self-driving vehicles and other autonomous machines that perceive and manipulate objects while respecting the laws of physics. "The next wave of AI is physical AI," Nvidia CEO Jensen Huang has said, describing systems that can "work among us" and learn how to perceive the world.
Who is building it?
Startups and established tech firms are investing heavily. Enactic's OpenArm training rigs are used by Nvidia and by universities such as Stanford to collect human demonstrations. Chinese firms including XPeng, Unitree Robotics and EngineAI are pushing humanoid and legged designs, while Boston Dynamics and others continue to demonstrate nimble, mobile robots.
'Any human role'
At a recent Guangzhou launch, XPeng unveiled a humanoid that walked and danced on stage. XPeng CEO He Xiaopeng has suggested robot volumes could one day rival car production, and XPeng co-president Brian Gu told AFP that with enough data and training humanoid AI could perform "almost any human role" — from nanny to home cook. Yet companies caution that fine manipulation and durable, replaceable components (for example, robot hands used in heavy lifting) remain costly.
On-the-job training
Unlike text models that learn from documents, physical AI must master vision and spatial relationships. For now, teleoperation — humans controlling robots to demonstrate tasks — is the most reliable way to collect training data. Enactic says 30–50 demonstrations per task can often be enough to fine-tune "vision-language-action" models. The plan is to deploy teleoperated robots in care homes to handle routine chores while collecting data that will eventually enable autonomous operation.
Real-world limits and mishaps
Progress is uneven. Videos show promising demos, but there have been embarrassing moments: a newly unveiled Russian humanoid stumbled and fell, and a US–Norwegian home assistant, NEO (priced at $20,000), has struggled with simple chores such as closing a dishwasher door even when teleoperated. Sara Adela Abad Guaman, assistant professor in robotics at University College London, warns of "a big gap" between AI software and the physical capabilities of machines. She notes that biological bodies are evolved to suit environments — and our sense of touch remains hard to replicate.
Industry momentum and concerns
Investment and acquisition continue. Morgan Stanley projects more than a billion humanoid robots could exist by 2050, and Japan's SoftBank recently called physical AI its "next frontier" as it agreed to buy ABB Robotics for $5.4 billion. But the field raises questions about job displacement, privacy, safety and whether current hype will translate into widespread, practical use.
Bottom line
Physical AI is moving from lab demos to pilot deployments — especially via teleoperation and targeted use cases such as care homes — but substantial engineering, cost and safety challenges remain before humanoid robots are common household helpers. The next few years will show whether these systems can reliably translate impressive demonstrations into everyday value.