For example, Siemens’ SIMATIC Robot Pick AI extends this vision of adaptability, transforming standard industrial robots once limited to rigorous, repetitive tasks into complex machines. Trained on synthetic data – virtual simulations of shapes, materials and environments – AI prepares robots to handle unpredictable tasks, such as picking an unknown item out of a chaotic bin, with greater than 98% accuracy. When mistakes are made, the system learns and improves through real-world feedback. Crucially, this is not something that can be solved with just one robot. Software updates will extend across the entire fleet, upgrading robots to operate more flexibly and meet the growing demand for adaptive production.
Another example is ANYbotics, a robotics company that creates 3D models of industrial environments that act as digital twins of the real world. Operational data such as temperature, pressure and flow are integrated to create a virtual replica of the physical facility on which the robot can train. For example, an energy plant can use site planning to create a simulation of the inspection tasks that robots will need to perform at the facility. This speeds the training and deployment of robots and allows them to be performed successfully with minimal field setup.
Simulation also allows you to scale up your educational robot fleet at virtually no cost. “Simulation allows us to create thousands of virtual robots to practice tasks and optimize their movements. This allows us to reduce training times and share knowledge between robots,” says Péter Fankhauser, CEO and co-founder of ANYbotics.
Because robots need to understand their environment regardless of orientation or lighting, ANYbotics and partner Digica created a way to generate thousands of synthetic images to train robots. By eliminating the arduous task of collecting huge numbers of real images from the shop floor, the time needed to teach robots what they need to know is drastically reduced.
Likewise, Siemens leverages synthetic data to create simulation environments to digitally train and validate AI models before deploying them in real products. “We use synthetic data to create variations in object orientation, lighting and other factors to help the AI adapt well to different conditions,” says Vincenzo De Paola, project lead at Siemens. “We simulate everything from the orientation of the pieces to lighting conditions and shadows. This allows models to train in a variety of scenarios, increasing their ability to adapt and respond accurately in the real world.”
Digital twins and synthetic data have proven to be powerful antidotes to data scarcity and expensive robot training. Training robots in artificial environments can quickly and inexpensively prepare them for a variety of visual possibilities and scenarios they might encounter in the real world. “We validate our models in this simulation environment before deploying them physically,” says De Paola. “This approach allows us to identify potential problems early and improve the model with minimal cost and time.”
The impact of this technology can extend beyond initial robotics training. By using the robot’s actual performance data to update the digital twin and analyze potential optimizations, you can create a dynamic improvement cycle that systematically improves the robot’s learning, functionality, and performance over time.
A well-trained robot at work
Organizations can benefit as AI and simulation usher in a new era of robotic training. Digital twins allow companies to deploy advanced robotics while dramatically reducing setup times, and the increased adaptability of AI-based vision systems allows companies to more easily change product lines in response to changing market demands.
