Shake-and-bake synthetic AI Vision Training data—just add CAD.
The Runaway Growth of Artificial Intelligence and Machine Vision
Artificial intelligence (AI) is currently experiencing a period of wild growth, with no end in immediate sight. Grandview Research estimated the 2021 AI market at $93.5 billion, and the forecasted growth could see a $1,811.8 billion market by 2030. This growth comes with a set of challenges, with one of the major ones being data. AI platforms need large groups of data to build decision-making algorithms, and that data needs to have both quantity and quality.

Inside the big umbrella of AI, vision systems are just as data dependent and positioned to make big growth gains. A recent Markets and Markets report estimates that the $15.9 billion spent worldwide on AI vision systems in 2021 could blow up to $51.3 billion by 2026. The report said that manufacturing companies will invest more in vision systems for component manufacturing, automation and quality control. Siemens is targeting this large and growing market and its training data challenge by using SynthAI. The new tool will generate training data from everyday CAD files.
In an industrial robotics setting, thousands of images are required to build a library that can teach the vision systems to learn. These images require the physical labor to set up a camera, monitor the setup, and then collect and sort the data. Traditional machine learning then requires that these images be annotated, using boundary boxes or contour outlines, to show the system the physical components that are being observed. Engineers then build code that will let the AI make decisions on its own, based on the data, to determine which parts are good and which are bad.
The Rise of Synthetic Data
Zachi Mann from Siemens says that synthetic data can help with two of the big roadblocks in the vision training process: data collection and annotation. Data collection can be difficult in a situation where only a few prototype components exist, or when long cycle times mean that only a few components can be studied every hour.
Creating synthetic datasets gives the engineers the ability to have a large number of data points that train the system to see the parts built during a typical day of production as well as expose the vision system to more special cases. Over a year of production, any vision system will see parts fabricated incorrectly or incompletely. Instead of waiting a full year before the system is fully trained or requiring physical labor to make a set of bad parts, synthetic data can generate a full range of parts from pristine to highly damaged.

Annotation is the other bottleneck that synthetic data can help to alleviate. Manually annotating datasets is tedious and repetitive work. Using a human to annotate this data can introduce human errors during a manufacturing shift. Synthetic data and annotation, however, can be more accurate because the system generated the data and will automatically recognize the differences present in the image data.
The next step in integrating this synthetic data with reality can take a variety of approaches, ranging from what Mann calls close to real and domain randomization. Close to real, true to its name, sees programmers spending their efforts trying to make the data as close as possible to reality. Different lighting conditions, the parameters of a specific vision camera, material properties of the components and visual background noise are all used to make the data seem as true to life as possible. This method creates more accurate trained systems when working in similar environments but is effort intensive.
Domain randomization takes the synthetically generated data and randomizes other parameters to create even more variation. Lighting, material properties of the component, and background noise can all be varied to create more data points that will improve the vision system’s learning. The idea is that with more alterations happening around the component, the system will learn to filter out these distractions and focus on the most important aspects of the operation. Domain randomization takes less effort for an engineer to automate and is more robust to changes in the system. However, domain randomization requires more data so the system can acclimate to a wider range of scenarios. This technique can also fail to meet requirements and requires some fine-tuning adjustments.
SynthAI Starts with a CAD Model
SynthAI seeks to take the complexity out of the process of generating quality synthetic data. The first step is uploading a CAD model into the SynthAI interface. The software then generates thousands of images with randomized parameters in—what Siemens says will be—a matter of minutes.
A machine learning model is also built and trained for the user to deploy with a physical set of components. Users can then download the model and dataset to train their own system and/or test the models. That workflow of uploading a CAD model to get thousands of test images and a trained model is a huge part of the machine vision training process.

This trained model comes with a Python environment setup and sample code for the user to test. If this setup doesn’t perform with real-world components as well as anticipated, the user can add images of the physical components into the datasets and strengthen the trained model. Users can also perform the annotation on images before adding them to the system, including boundary boxes or contour shapes into the images. This fine-tuning of the system will help bridge the gap between the synthetic data and reality.
The process takes place entirely through a website interface and does not require the user to have specialized software or equipment. Siemens has been working on the SynthAI project for a while and already has projects completed with quality inspection, flexible robotic assembly, pick and place, and sorting and kitting applications.
What Does It All Mean?
AI is going to transform our lives in ways that we can’t even imagine yet, and it’s fascinating to watch the origins happen in manufacturing. Data will continue to be an issue, and even with the big cloud companies all working on ways to deliver more data, it still feels like more data will always be needed. Synthetic data for training AI isn’t necessarily new, but the approach that Siemens is taking here to generate the data and build training models feels like a jump in technology.
AI has a Wild West feeling to it right now, as standards are being built on the fly and demand continues to grow exponentially. SynthAI reflects that fast-moving progress, described by Siemens as a “solution still in the making,” with features and capabilities being added often.
The training model created by the software might not be completely accurate to reality, but the user can add images and better train the model. It has the feel of a collaborative environment where the customer is taking on some of the responsibility to provide the solution, knowing that the extra work will create a superior product in the end.
Operating using SynthAI can now help a company to develop a robotic arm, for instance, and be training that arm and vision system almost at the ground floor. Cutting the image generation and annotation stages out of the vision system training process is big. Instead of waiting for prototypes of the robotics system and prototypes of the components that will be manipulated by the robot, synthetic data can start training the system almost as quickly as the system can be designed. It will be interesting to see what SynthAI looks like in a year or two and what it might be doing then.