ABB ROBOTICS INDUSTRIALIZES PHYSICAL AI.
ABB Robotics is turning the next frontier of artificial intelligence into industrial reality. Combining more than 50 years of robotics leadership with advanced AI technologies, industrial engineering expertise and a global ecosystem, we are redefining what robots can do — and what industries can automate.
We make Physical AI productive, reliable and scalable, bringing a new level of intelligence, autonomy and versatility to industrial operations without compromising the precision, safety and performance production demands.
AUTOMATE MORE. DEPLOY FASTER.
- Industrial Physical AI is Physical AI built for real production: the volume, precision, reliability and efficiency manufacturing demands.
- Robots run on trained AI models instead of being programmed movement by movement.
- ABB Robotics’ Physical AI Toolchain creates training data, trains AI models and runs them on the robot, backed by a virtual controller that runs the same firmware as the physical robot.
- With Fashion Inductor, AI Robot Trainer can reduce engineering and commissioning effort by up to 75%, making it faster and easier to train and deploy robots for new tasks.
- Backed by more than 50 years of robotics and Connected Care, which keeps systems performing long after deployment.
UNDERSTANDING PHYSICAL AI
<p>Physical AI is expanding the boundaries of automation, enabling robots to take on tasks previously considered too complex, variable or uneconomical.</p>
What is Industrial Physical AI?
Industrial Physical AI brings artificial intelligence into real production environments, where reliability, precision, safety and performance are essential. ABB Robotics turns these advances into industrial performance at scale.
Unlike isolated demonstrations in controlled conditions, a robot on a production line must perform at cycle time, handle variation, operate safely around people and keep working for years. It must be secure, supportable and integrated with the systems already running the operation.
That is what ABB Robotics does: we take Physical AI from the lab to industrial scale.
From Industrial Physical AI to Autonomous Versatile Robotics
Industrial Physical AI is changing what robots themselves can do. By combining intelligence with greater autonomy and versatility, robots can move beyond fixed programming to adapt to changing environments and perform a wider variety of tasks.
We call this Autonomous Versatile Robotics (AVR™): ABB Robotics’ vision for the next generation of intelligent industrial automation, bringing together sensing, reasoning, learning and action with the precision, reliability and safety industry demands.
What is Physical AI?
For decades, industrial robots have excelled at specific jobs in predictable environments, following predefined instructions.
Physical AI changes that.
It combines AI reasoning with physical robotics – the sensors, control systems and mechanics that let a robot act in the world.
It fundamentally changes what robots do, where they operate and the value they create – a step change for advanced manufacturing.
Instead of programming every movement of every robot, engineers train the AI models that guide them.
These models learn from examples to recognize what is in front of them and decide what to do next. When the task or conditions change, the model can be retrained rather than the robots reprogrammed.
What makes Physical AI possible today?
Teaching machines from examples is not a new idea. Until recently, doing it at industrial scale was not practical, because every task needed huge amounts of data and had to be learned from scratch.
Three recent advances have changed that:
- Foundation models give AI a broad starting knowledge of images and language, so a model no longer starts from scratch.
- Vision-language-action models and world models connect what a robot perceives with how it acts, helping it understand its environment, choose an action and anticipate what happens next.
- Simulation is now accurate enough to train in. Models can practice thousands of variations before a robot touches a real part, improving through trial and feedback with reinforcement learning.
Five million industrial robots now work in factories worldwide, almost all of them on fixed, repeating tasks, according to the International Federation of Robotics. Physical AI opens the door to work that has historically been too variable to automate and to industries automation has not reached.
A new generation of robots is emerging – more intelligent, more versatile and more capable of doing more things, in more places.
Physical AI vs. Traditional Programming
| Traditional robot programming | Physical AI | |
| How the robot is set up | Engineers program every movement | Engineers train the AI models that decide how the robot handles each task |
| When the task changes | Robots are reprogrammed | The model is retrained |
| Best suited to | Specific jobs in predictable environments | Work that varies from one item or situation to the next |
HOW IT WORKS: FROM TRAINING TO PRODUCTION
Designed for garment picking in logistics, Fashion Inductor handles items that almost never arrive the same way twice: crumpled, folded or overlapping, in different fabrics, shapes and colours the system has never encountered before. It is exactly the kind of variation that conventional robot programming struggles to handle. Fashion Inductor already sorts up to 1,300 items per hour per cell, with picking accuracy targets above 99.5%.
-
Start from a pretrained model
There is no need to train from scratch. Fashion Inductor starts from an AI model already pretrained on garments, giving it relevant knowledge before customer-specific training begins.
-
Fine-tune it on the customer’s items
Using AI Robot Trainer, the model is fine-tuned in the cloud on images from the customer’s own operation. Because the model already understands the domain, around 50 images can be enough to reach production quality, compared with thousands for a model with no prior knowledge of garments.
-
Deploy
The fine-tuned model becomes a deployable file, uploaded to the robotic cell as a new version of the system’s intelligence. No permanent cloud connection is required.
-
Improve
The system monitors performance, detects when it starts to drift from its targets and recommends retraining the AI model. New product families or changes in lighting are no longer a challenge. Improvement is controlled and deliberate, giving integrators and local operators full visibility, unlike black-box AI.
-
The result
Engineering and commissioning effort is reduced by up to 75%, and an AI model can be retrained and put into production in minutes. Tasks never considered become worth automating.
ABB ROBOTICS’ PHYSICAL AI TOOLCHAIN
<p>Our integrated Physical AI Toolchain connects simulation, AI training and robotic execution, enabling intelligent automation to move from development into industrial deployment.</p>
Real images can be difficult, expensive or impractical to collect at scale. Powered by NVIDIA Omniverse, RobotStudio HyperReality generates synthetic training data from highly realistic simulations, allowing models to experience variations that would be difficult to capture physically.
The robot in simulation is controlled exactly as the physical robot, thanks to ABB Robotics’ unique virtual controller that shares the same firmware. As announced with NVIDIA, this enables up to 99% accuracy from virtual training to real-world deployment.
AI Robot Trainer fine-tunes pretrained AI models in the cloud, using images from the customer’s own operation. The result is a deployable AI model file.
EyeMotion gives a robot the perception and control to act on what it sees. It captures images from the operation for training, and runs the trained model on the robot to recognise objects and adjust to what is in front of it.
RobotStudio will support the Model Context Protocol (MCP), enabling integrators and partners to connect their own AI agents. These agents can take on engineering tasks such as converting CAD files for simulation or retrieving robotic cell libraries, helping engineers move faster from concept to deployment.
BUILDING PHYSICAL AI AT SCALE
Physical AI takes an ecosystem
- Physical AI takes an ecosystem
- Connected Care: built to keep performing
- Building what comes next
PHYSICAL AI TAKES AN ECOSYSTEM
Physical AI innovation needs collaboration in AI models, compute, simulation, sensing and robotics, and the expertise to make them work together outside the lab.
Our collaboration with NVIDIA underpins RobotStudio® HyperReality. We work with Landing AI on the vision models behind our first AI-trained applications, and with PSYONIC on dexterous robotic hands.
System integrators are equally important. They turn these technologies into working applications, operation by operation. Our tools therefore need to work as effectively in their hands as they do in ours.
Supporting the entire ecosystem is ABB Robotics’ global reach: our footprint, channel network and service organisation help support robotic systems wherever they are deployed.
Together with our technology partners and our system integrators, we take Physical AI from individual applications to industrial scale.
Through ABB Robotics Ventures, we also collaborate with and invest in early-stage companies shaping the future of Physical AI and Autonomous Versatile Robotics (AVR™), turning new ideas into real-world impact for our customers.
CONNECTED CARE: BUILT TO KEEP PERFORMING
Getting a robot into production is only the beginning. It is expected to perform for years.
Connected Care brings together connectivity, software, data, AI and ABB’s lifecycle expertise to keep it performing.
Connected robots share data from operation, so issues can be detected earlier and diagnosed remotely, often before they stop production. Service moves from reacting to problems to anticipating them.
That support runs through predictive and condition-based maintenance, remote diagnostics, spare parts and field service, and modernisation as operational needs evolve.
It is what makes Physical AI a long-term production asset rather than a one-off deployment.
BUILDING WHAT COMES NEXT
For more than 50 years, robots have transformed the work that could be programmed. Physical AI expands that frontier to work that could not. That means more tasks, more applications and more industries where automation can create value.
ABB Robotics is building the technology, engineering expertise and ecosystem to make Physical AI work where it matters: in real operations, every shift, for years.
FAQ
Physical AI combines AI reasoning with physical robotics – the sensors, control systems and mechanics that let a robot act in the real world. Instead of following fixed instructions, a physical AI robot uses trained AI models to recognise what is in front of it and decide what to do next.
Industrial physical AI is physical AI built for the volume, precision, reliability and efficiency that real manufacturing demands. It performs at production cycle time, handles variation, operates safely around people and is secure, supportable and integrated with existing operations.
Traditional industrial robots follow predefined instructions for specific jobs in predictable environments. With physical AI, engineers train the AI models that guide the robots instead of programming every movement, and when tasks or conditions change the model can be retrained rather than the robots reprogrammed.
Training starts from a pretrained AI model, which is fine-tuned on images from the customer’s own operation, deployed to the robotic cell as a model file, and improved by retraining when products, materials or conditions change. Because the model already has relevant knowledge, far less data is needed than when training from scratch.
The Physical AI Toolchain is a set of ABB Robotics software tools that create training data, train AI models and run them on the robot. RobotStudio HyperReality generates synthetic training data, AI Robot Trainer fine-tunes the model, and OmniCore EyeMotion runs it on the robot.
RobotStudio HyperReality is ABB Robotics’ simulation capability, powered by NVIDIA Omniverse, that generates synthetic training data from highly realistic simulations. The simulated robot is controlled exactly like the physical one, because ABB’s virtual controller shares the same firmware.
No. Models are fine-tuned in the cloud, but the result is a deployable file uploaded to the robotic cell as a new version of the system’s intelligence, so no permanent cloud connection is required to run it.
ABB Robotics’ Connected Care combines connectivity, software, data, AI and lifecycle expertise. Connected robots share operating data so issues can be detected earlier and diagnosed remotely, supported by predictive and condition-based maintenance, spare parts, field service and modernisation.