.png){kind=logo}
Physical AI is helping robots understand environments, instructions and complex physical tasks.
Google DeepMind and NVIDIA are developing technologies powering next-generation intelligent robotic systems.
Humanoid robots face challenges involving safety, reliability, costs and widespread commercial deployment.
Artificial intelligence is moving beyond screens and software into the physical world. Robots that once depended on tightly programmed instructions are increasingly being equipped with AI systems that can interpret their surroundings, understand instructions, plan tasks, and respond to changes in real time.
This shift is driving the rise of what the industry calls physical AI – systems that connect artificial intelligence with machines capable of sensing and acting in the real world. Google DeepMind, NVIDIA and a growing group of robotics companies are developing technologies aimed at making robots more adaptable, rather than limiting them to repetitive, predefined tasks.
Traditional industrial robots have been highly effective at repetitive work. But their usefulness is often limited when the environment changes or when a task requires judgment.
AI is beginning to address that gap. Modern robotics models combine computer vision, language understanding, spatial reasoning and motor control, allowing machines to process information from cameras and other sensors before deciding what to do.
Google DeepMind’s Gemini Robotics models illustrate this transition. Gemini Robotics 1.5 is designed as a vision-language-action model that can turn visual information and instructions into motor commands. Its capabilities include understanding physical environments, adapting behavior to new situations, and responding to everyday commands.
In April 2026, Google DeepMind introduced Gemini Robotics-ER 1.6, a reasoning-first model designed to improve how robots understand the physical world.
The model focuses on visual and spatial understanding, task planning, and success detection. It can also interpret complex instruments, including gauges, helping robots deal with situations that require more than simple object recognition. Google said the model was developed to give physical agents greater autonomy through enhanced spatial reasoning and multi-view understanding.
The significance is broader than a single model. Robots need to know not only what is in front of them but also what those objects mean in context, what action should come next, and whether the task has actually been completed.
NVIDIA is taking a different but complementary approach by building the infrastructure needed to train and deploy intelligent robots. The company has brought together technologies including its Cosmos world models, Isaac simulation frameworks and Isaac GR00T models to support robotics development. In March, NVIDIA said robotics companies across industrial, surgical and humanoid applications were using its technology to develop physical AI at scale.
In July, NVIDIA expanded its robotics push with new Jetson Thor computers designed to run AI models at the edge. The company said general-purpose robots and autonomous machines are moving from research environments towards real-world deployment, increasing demand for compact and power-efficient computing systems.
Greater autonomy also creates a more difficult safety problem. A robot operating independently in a factory, warehouse, or hospital cannot be treated like a software application.
NVIDIA announced Halos for Robotics in June 2026, describing it as a full-stack safety system for physical AI. The system brings together AI computing, sensor connectivity, software and safety functions into a common architecture. Agility, a humanoid robotics company, became the first company to work with NVIDIA on incorporating elements of the system into its own safety framework.
The focus on safety reflects a larger challenge for the sector: a robot must be capable of acting independently while remaining predictable when conditions change.
Humanoid robots have become one of the most visible areas of the AI robotics race. China, the US and other technology hubs are investing heavily in machines designed to work in environments built for humans.
China’s Unitree has emerged as a prominent example. The company, founded in 2016, is preparing to become the first humanoid robot manufacturer to list on China’s mainland stock market. Its robots, including the G1, H1 and R1, have gained attention for capabilities such as running, dancing and performing martial arts. Reuters reported that more than 40% of Unitree’s sales in 2025 came from overseas.
The broader market, however, is still at an early stage. Commercial adoption remains concentrated largely in research and education, while companies continue to work on lowering costs and improving reliability.
AI has made robots better at perception, reasoning and adaptation, but that does not automatically make them commercially viable. Robots still need reliable hardware, large amounts of training data, powerful computing, and safe control systems. They also have to operate consistently in environments that are far less predictable than digital applications.
That is why the next phase of AI in robotics will be judged less by demonstrations and more by deployment. The industry has already shown that AI can help machines see, understand and act. The harder question is whether those capabilities can be delivered reliably, safely and cheaply enough for widespread use.
For now, physical AI is moving robotics in that direction. The technology is shifting the focus from robots that execute instructions to machines that can interpret situations and make decisions within the physical world.
