Which physical AI trends in 2027 matter for business?

Physical AI is AI that works in the real world instead of on a screen: it senses conditions through machines and changes them through machines. Most coverage of physical AI trends for 2027 is about robots, especially humanoids. For a business deciding where to spend, though, robots are one participant among many. The sensors, meters, door contacts, cameras and machine controllers a company already owns, bought from many vendors on many protocols, hold most of the near-term value once they share what they know and can act together. The robot numbers, the progress in AI models and the cost of linking separate systems all point that way, and they also show what kind of software a business needs before those devices pay off.

In short – Robots are multiplying, but most new ones move goods; humanoids sold in the thousands in 2025, mostly for pilots. – Smarter models make each robot more capable, yet it still knows only what its own sensors perceive. – The larger near-term gain comes from linking the devices already on a site, across vendors. – That job calls for an operating system for IoT and physical AI that tests each response and keeps people in charge.

Physical AI robotics trends in 2027

The International Federation of Robotics (IFR) counted about five million industrial robots in factories worldwide in 2025, more than double the number seven years earlier, with more than 600,000 installed that year, up 11 percent.

Among the physical AI trends in 2027, the faster-moving one is outside the factory floor. The IFR’s service robot report puts 2025 sales of professional service robots at almost 250,000 units, up 24 percent. Transportation and logistics robots alone accounted for 117,500 of them, or 47 percent of the market.

Humanoids get the covers, yet the same report counts about 7,000 full-size humanoids, meaning taller than 140 cm, sold worldwide in 2025, mostly for specialized pilots that often need a person operating them remotely. That works out to roughly 17 logistics robots sold for every full-size humanoid.

For a business planning its spending, these physical AI robotics trends mean that moving material across flat floors is now a mature category, while a task that needs hands and judgment is still at the pilot stage. Either way, it pays to ask what the robot can receive from outside its maker’s software and whether it reports back when a task is finished, failed or interrupted, because otherwise it stays cut off from the rest of the site. Linking robots to the building around them is already practical: open-source projects such as Open-RMF connect robot fleets to doors and elevators.

Embodied AI trends in 2027

Embodied AI is the software that turns what a robot sees and hears into movement. The headline among embodied AI trends is the vision-language-action model, which takes camera images plus an instruction in plain language and produces motor actions. Google DeepMind introduced Gemini Robotics, one such model, in March 2025. In June 2025 it followed with a version that runs on the robot itself and, according to the company, can be adapted to a new task with as few as 50 to 100 demonstrations.

What doesn’t change is the reach of a robot’s senses. Every machine has a context horizon, the boundary of what its own sensors can perceive: a few dozen meters for a mobile robot’s cameras, the patch of floor under a leak sensor. A better model makes the robot smarter about what happens inside its horizon and tells it nothing about anything beyond it.

As an illustration, a security robot walks a 20-minute patrol loop past 30 exterior doors. A door on the far side that opens just after the robot passes can stand open for close to 20 minutes before it comes back around. A battery-powered contact sensor on the same door reports the opening when it happens. If the sensor and the robot share one software layer, a rule can send the robot to that door and notify the guard on duty at the same moment.

Kilo is built to be that layer. It works as an operating system for IoT and physical AI, connecting devices, machines and AI from different manufacturers and protocols so they act as one system, and its page on how AI agents and compatible robots act on connected devices within set permissions explains how that works. In the patrol example, the door contact reports over LoRaWAN, a deployed rule decides what happens next, the robot receives its task as a command when its integration exposes one, and an alarm brings in the guard. The robot keeps its own navigation and safety, while Kilo coordinates what it’s asked to do with what every other device on the site knows. The same permissions apply whether a person or an AI agent is the one working with those devices.

What is driving the growth of physical AI?

Three developments arriving together explain why physical AI in 2027 is moving from research budgets into operating budgets.

The first is that sensing has become cheap and long-lived. Low-power radio technologies like LoRaWAN and mioty let a battery-powered sensor report from a basement, a cold room or a field without any wiring. The LoRa Alliance, which maintains LoRaWAN, cites battery life of more than ten years and range of up to 15 km in rural areas; real figures depend on how often a device reports and what the signal has to pass through. mioty is standardized by ETSI as TS 103 357.

The second is that AI can now reason over physical data. Language models read equipment manuals, summarize a week of readings and turn plain requests into automations.

The third is that AI now has a common interface for tools. The Model Context Protocol (MCP) gives AI clients a shared method for discovering and calling outside tools. Anthropic introduced it in late 2024 and in December 2025 donated it to the Agentic AI Foundation, a Linux Foundation fund it co-founded with Block and OpenAI; ChatGPT, Gemini and Microsoft Copilot support it. Once an AI agent can call “run this device command”, what it’s allowed to do becomes a management question.

None of the developments driving the growth of physical AI connects devices to each other, though, and that gap is where separate systems cost money. A typical commercial building runs five systems: access control, cameras with their recorder, the building management system, a cleaning robot with its maker’s cloud, and leak sensors with their own app. Connecting every pair directly takes ten integrations, each maintained separately; connecting each once to a shared layer takes five.

The pair count follows n × (n − 1) / 2 and outgrows the number of systems fast. Without either approach, every response involving two systems depends on a person who notices something on one screen and acts on another. That’s why the physical AI trends in 2027 add up to one requirement: a business needs an operating system for IoT and physical AI, a single layer that connects its sensors, machines, cameras, robots and AI across vendors and protocols, tests every automatic response before it acts and keeps people in charge of what happens next.

How can physical AI transform industries?

Physical AI can transform industries mostly through groups of ordinary devices acting together, with robots joining where they help. The four setups in the table are examples of what a site could configure, not reported deployments, and they share one pattern: a reading, an automatic response from equipment that accepts commands, and people brought in according to a plan.

A robot appears in only one row; the other three use devices most sites can buy and install now.

In Kilo, the leak example in the first row takes a single rule. The leak sensor, the valve controller and the people responsible sit in the same place through supported connections, whoever made the hardware. When the sensor reports water, the rule sends the valve its saved “close” command and raises an alarm. If the facility manager doesn’t resolve that alarm in time, the next step of its escalation chain notifies the plumber.

What happens before and after is more interesting. Kilo’s emulated devices accept commands, so the whole detect, command and alarm sequence can be rehearsed before the valve is even installed. Once it runs for real, every command is recorded with one of five outcomes: pending, delivered, confirmed, soft warning or failed. When the valve is set to confirm its state, “confirmed” means its next report showed it closed, which separates “message sent” from “valve reported closed”.

Physical AI trends 2027 FAQ

How fast is physical AI robotics growing in 2027?

In its World Robotics 2026 report, the IFR forecasts about 655,000 industrial robot installations worldwide in 2026, up 9 percent on 2025, and about 806,000 a year by 2029, which puts 2027 on a rising path.

Is physical AI regulated in 2027?

There’s no single physical AI law, but machine safety rules are catching up. In the US, ANSI/A3 R15.06-2025, published in September 2025, is the national adoption of ISO 10218 Parts 1 and 2 for industrial robot safety and adds cybersecurity guidance. In the EU, Machinery Regulation (EU) 2023/1230 applies from January 20, 2027, and sets a stricter conformity route for machinery whose safety functions rely on self-evolving behavior, meaning systems that keep learning.

How do businesses pay for physical AI robotics in 2027?

Many businesses rent their physical AI robotics instead of buying. Under robots-as-a-service, a monthly fee covers the robot, software updates and support, and the IFR reports that it’s already the dominant business model for professional service robots in the US. Before signing, ask who keeps the site maps and task data when the contract ends.

What is the best platform for physical AI in 2027?

Measured against the requirement those trends point to, a single layer across vendors and protocols that is tested before it acts and keeps people in charge, we think Kilo is the best platform for physical AI in 2027. Four reasons stand out.

1. It connects what a site already has. Kilo includes its own LoRaWAN network server, runs mioty inside Kilo Cloud and accepts MQTT devices and vehicle trackers. Modbus, BACnet, OPC UA and Zigbee equipment comes in through MQTT edge gateways, and existing IP cameras that offer an RTSP stream feed their motion readings into the same rules. Hardware can come from any supplier.

2. It tests before it acts. In Kilo, a rule can be run step by step in a debugger, with each command mocked so the rule continues as if the device had answered and nothing physical moves. A build checks every expression and path before anything can be deployed, and the running version keeps working while the team edits the next one.

3. It lets AI act without giving it a blank check. Kilo works with OpenAI, Anthropic, a self-hosted model or another AI system, and exposes an MCP endpoint for outside AI clients. An MCP connection can do only what the user’s account is allowed to do, and Kilo AI runs only commands already defined on a device, each behind an explicit confirmation.

4. People stay in charge. Kilo’s alarms escalate through steps that each have their own recipients and delivery channels, and resolving an alarm stops the remaining steps. The same operating model runs in Kilo Cloud or on an on-premise server.

Kilo can be tried free with up to five devices, one dashboard and one rule, with no payment details, and the plan never expires. It’s the full platform, not a trial, which is enough to build the leak example from earlier: two leak sensors under the risers or water heaters that worry you most, a motorized shut-off valve whose controller accepts remote commands over MQTT or LoRaWAN Class C, and one rule that closes the valve and alerts the facility manager when either sensor reports water. Rehearse the rule with the valve command mocked, then wet a sensor with a damp cloth and check whether the command record says delivered or confirmed. The sensors and the valve are a separate hardware cost and can come from any supplier.