AI is leaving the screen. Here's what happens when it gets a body.

Most of the AI conversation over the last few years has stayed on the screen. It has been about models, chips and chatbots, the cognitive side of the technology. Barclays Research captures the turn in a single line of its Impact Series report, The Future of Work: AI Gets Physical: AI-powered humanoid robots are stepping out of the research lab and into the economy. The same shift is happening well beyond humanoids, in warehouses, factories, hospitals and fields, wherever a machine now needs to see what is in front of it and decide what to do. That is what people mean when they say physical AI.

01Teaching a robot to move

One of the biggest fields in robotics is also the one most people overlook, which is movement. A robot that cannot reliably get from one place to another, or keep its balance on an uneven floor, cannot do much else. Traditionally that meant engineers hand-building a model of the robot and its surroundings and tuning it for each situation. The newer approach is to let AI learn the movement instead.

On the navigation side, robots use techniques such as SLAM to build a map of an unfamiliar space while tracking where they are inside it, then plan routes around obstacles and reroute when a corridor is blocked. Warehouse robots do this all day, moving through aisles and carrying goods without a human steering them. On the locomotion side, reinforcement learning lets a robot practise a gait thousands of times in simulation, being rewarded when it stays upright and moves well, until the behaviour carries over to the real machine.

Harvard researchers have shown how far this has come. Yilun Du's team at the John A. Paulson School of Engineering and Applied Sciences presented Diffusion-MPC at the IEEE International Conference on Robotics and Automation in June 2026, a control system for four-legged robots that combines model predictive control with a generative diffusion model, the kind of AI more often associated with making images. The robots could adapt to new tasks and terrain without being retrained, adjusting to different speeds, directions and gaits while staying within physical limits and safety rules. Du described it as a step toward general-purpose locomotion controllers.

02Giving it eyes, and a way to understand what it sees

Vision models are the other big reason robots are suddenly capable of more. Computer vision has been part of robotics for a long time, but today's models are fast and accurate enough to run on the robot itself. Object detection models such as YOLO can pick out and classify several things in a scene at once, segmentation models label every pixel so the robot understands the shape of the space, and tracking models give each object an identity so the robot can follow it over time. Cameras are often combined with LiDAR, radar and other sensors, and the robot fuses them into one picture of its surroundings.

The more recent change is that vision and language have been brought together. Vision-language-action models take a camera image and a plain instruction, such as picking up the red block and placing it in the bin, and turn them straight into motor commands. Google DeepMind's RT-2 transferred knowledge learned from web-scale images and text into robot control, Figure's Helix uses the same approach to let a humanoid see, interpret an instruction and plan a physical action without a pre-programmed routine, and NVIDIA's GR00T gives humanoids a foundation model that learns from demonstration, simulation and video. The result is robots that can handle more complicated work, at higher speed and with greater accuracy, because the model understands the scene instead of following a fixed script.

03Where AI is used in a robot

Movement and vision are the most common use cases right now, but AI runs through almost every layer of a modern robot. It helps to see them laid out together.

Much of this runs on the robot itself, at the edge, where speed matters for things like obstacle avoidance. The heavier work of training models on large datasets happens in the cloud. Between the two, a robot is increasingly a node in a larger system, rather than a self-contained machine.

04A perfect storm for physical AI

Several forces that used to move separately are now landing at the same time. The first is the cost of building the hardware that collects data. Customised machines for sensing the real world are far easier and less expensive to put together than they used to be, in large part because of vibe coding. One maker recently described using Claude with the Atopile design tool to produce a design for an ESP32-S3 development board from a detailed prompt, approving each step instead of writing the code himself. He was careful to say that the board had not yet been manufactured or tested, and that AI is not ready to replace a skilled engineer, but the distance between an idea for a sensing machine and a first prototype has clearly shrunk.

The second is connectivity. Wi-Fi and Bluetooth cover the room, 5G covers the city, and satellite links now reach the places those networks do not. 3GPP Release 17 standardised non-terrestrial networks so that ordinary NB-IoT devices can talk to satellites directly, and Nordic Semiconductor's nRF9151 chip has already completed chip-to-cloud communication through Sateliot's satellites. Satellite still costs more per device than terrestrial cellular, so for now it works best as a complement for coverage gaps, but taken together these technologies make always-on communication between machines possible from almost every corner of the planet.

The third is that the AI has become small enough to live on the device. Dell's Daniel Cummins has called 2026 the year of the small language model, pointing to Gartner's forecast that by 2027 organisations will use small, task-specific models three times more than general-purpose LLMs. Small models need far less compute and energy, and they can run locally without depending on a data centre. With smaller specialised models and small foundation models, a smart edge device powered by a modest battery can be assembled for a very specific use case with relatively little effort.

Cheap sensing, always-on connectivity and capable small models arriving together is, in our view, a perfect storm for physical AI to become the next AI boom, and the numbers are already moving. Barclays Research estimates the global humanoid robotics market at around 2 to 3 billion US dollars today, with the potential to reach 200 billion by 2035 under the most optimistic scenarios. It points to production costs that have fallen roughly thirtyfold over the past decade, to progress in AI software, physical motion and batteries, and to demographic pressure and labour shortages creating demand in manufacturing, logistics and healthcare. The International Federation of Robotics counts about 4.66 million industrial robots operating worldwide, and China alone has an operational stock of around 2 million in its manufacturing industry, roughly 4.5 times that of Japan, the next largest. It also accounts for 54 percent of annual industrial robot installations.

From warehouse to town-house, from peacetime to wartime, from highways to airways, physical AI in many forms is going to be built to carry out work that was never possible just a few years ago. Defence offers an early signal. On 30 September 2026, US Defense Secretary Pete Hegseth announced Project Meridian, a 120-day Pentagon study of the future of warfare that will recommend the capabilities the military should develop. It is co-directed by Elon Musk, Anduril co-founder Palmer Luckey and former House Speaker Newt Gingrich, which means two of its three leads come from companies that build physical, autonomous machines. It was announced alongside a new Autonomous Warfare Command focused on AI, robotics and biotechnology, with final findings due by 28 January 2027.

"The most useful robots will not be the ones that look most like people. They will be the ones plugged into the right data and the right workflows."

Singapore is building for this directly. In May 2026 the Infocomm Media Development Authority announced that the government and eight industry partners will research, test and deploy physical AI at Punggol Digital District, a living testbed for robotics and embodied AI due to launch later this year. Certis, DHL, Grab and QuikBot are the initial design partners, and the use cases include delivery, cleaning and security patrols that extend beyond office hours and into hard-to-reach spaces. IMDA describes robots working alongside human staff, with workers moving into supervision and operations management roles.

05What we tried: a patrol triggered from WhatsApp

Recently we ran a simple experiment of our own. We took robots equipped with Wi-Fi and connected them to our proprietary AI workflow platform. In a simulated environment, we then triggered a patrol with a message on WhatsApp. The robot acted as a surveillance and monitoring vehicle, moving through a facility to take pictures of the equipment and record temperature readings as it went.

What interested us most was where the intelligence sat. The robot did the moving and the capturing. Everything that turned a photograph and a number into a decision happened in the workflow, which is the same layer we already use to automate emails, invoices and approvals. The robot became one more way for the workflow to reach into the physical world and bring fresh information back, and the information arrived in a place where it could be analysed, compared against a threshold and acted on.

Looked at from a distance, the patrol followed a pattern we already know well. A machine takes in real-time inputs through sensors, decides what to do with them, and then takes the appropriate action. That is the same framework as our Trusted MultiVerse pipeline, which moves from Data to Knowledge to Action. In the patrol, the images and temperature readings were the data, the vision model's reading of what they showed was the knowledge, and the suggested action was the action. Trusted DataVerse, Trusted KnowledgeVerse and Trusted ActionVerse were built around documents and digital records, but the shape of the pipeline carries over to a machine that is sensing the physical world.

It also connects to an idea we wrote about in our last piece, on digital transformation. Data becomes far more useful when it is mixed with events happening in the outside world and kept up to date. A patrol robot is a simple way of producing exactly that kind of fresh, real-world data on demand, and a workflow that can read it gives a facility team something closer to a live view of their equipment than a periodic manual check.

06Where we go from here

This is only the start of our experimentation. We wanted to dip our toes into what physical AI makes possible, and a patrol that can be called up from a phone and returns with an assessment of the equipment is a small, concrete first step. Because it sits on top of our existing workflow platform, each thing we learn here can be reused elsewhere, whether that means other kinds of robots, other sensors, or other facilities.

We believe a platform that can address both digital and physical AI offers almost unlimited power to transform any industry, present or future. The same pipeline that turns a filing room into a conversational knowledge base and automated workflows can sit behind a robot, a sensor or an edge device, so the data a machine collects in the field arrives in a place where it can be understood and acted on.

As physical AI moves from research labs into everyday workplaces, we want to make sure it is introduced carefully and in a structured way, with a person deciding what to do when the system raises a flag. The robot can walk the floor and the model can read what it finds, and the people running the facility remain the ones who decide what happens next. That is how we intend to keep experimenting.