Physical AI: How Foundation Models Are Changing Robotics

For decades, industrial robots have been extremely precise and extremely narrow. A welding robot welds the same seam a million times. Change the part slightly and someone has to reprogram it. The new wave of robotics research is trying to change that, using the same ingredients that made language models general: large models, large and varied data, and lots of compute.

What "physical AI" means

Physical AI covers AI systems that sense and act in the real world. That includes:

• robotic arms in factories and labs,
• autonomous mobile robots in warehouses and hospitals,
• humanoid and legged robots,
• self-driving vehicles and delivery robots,
• drones and agricultural machines.

The defining challenge is that mistakes have physical consequences, and the real world is far messier than any dataset.

Robotics foundation models

Researchers and companies are training large models on robot data from many different machines and tasks. The leading approach is the vision-language-action (VLA) model: it sees through cameras, reads an instruction like "put the red mug in the dishwasher", and outputs motor commands.

What these models add:

• Generalization. Handling objects and layouts they were not specifically trained on.
• Natural-language tasking. Operators describe tasks rather than programming waypoints.
• Faster skill learning. New tasks from a modest number of demonstrations instead of months of engineering.

Major AI and chip companies, along with well-funded robotics startups, have released VLA models and robot learning platforms over the past two years, and progress is rapid.

Why simulation is central

Collecting robot data in the real world is slow, expensive and occasionally dangerous. Simulation lets teams generate millions of practice runs, vary lighting, objects and physics, and test edge cases safely. This is where game technology meets robotics: physics engines, realistic rendering and procedural content generation all come from game development. We cover this in robotics simulation with game engines.

Where robots work today

• Warehouses and fulfillment. Mobile robots move shelves and totes; arms pick items of increasing variety.
• Manufacturing. Collaborative robots work beside people on assembly, inspection and machine tending.
• Healthcare and labs. Lab automation, pharmacy dispensing, hospital logistics.
• Agriculture. Harvesting assistance, weeding, monitoring.
• Inspection. Legged robots and drones inspecting energy, industrial and construction sites.

What is still hard

• Reliability. A demo that works 90% of the time is impressive; a production line needs far higher.
• Dexterity. Handling soft, deformable or tangled objects remains difficult.
• Safety and certification. Robots near people must meet safety standards and regulations, including machinery rules in the EU and UK.
• Cost and maintenance. Hardware, integration and upkeep often outweigh the software cost.

How businesses should approach robotics now

1. Start with the process, not the robot. Identify repetitive, physically demanding or hazardous tasks in structured environments.
2. Consider proven systems first. Mobile robots, arms with vision and automated storage have strong track records.
3. Invest in software around the robot. Fleet management, integration with warehouse and ERP systems, monitoring and analytics decide whether robots deliver value.
4. Pilot with clear metrics: throughput, error rate, uptime, safety incidents, payback period.
5. Build simulation capability. A digital twin of your cell or facility lets you plan and test before buying hardware.

Physical AI is moving from research to deployment faster than most expected. Companies that build the data, simulation and integration foundations now will be best placed to adopt more capable robots as they arrive.