Automation has been a key part of manufacturing for over a century now, from the simple assembly lines of the past to the advanced, autonomous robotics of today. As the stresses placed on manufacturing systems continue to increase, however, the abilities of automated systems must increase as well. To meet the manufacturing demands of the 21st century, factory robotics must move beyond inflexible, hard-coded orders and gain the ability to quickly adapt to changing conditions — whether they be sudden business demands or new production requirements. This level of flexibility requires artificial intelligence (AI) certainly, but not just any AI; rather AI that can understand and interact with the real world. In other words, physical AI.
Physical AI refers to the application of artificial intelligence beyond the digital realm, where AI systems can perceive, reason and act in the physical world in real time. It combines advanced sensing technologies (such as vision, audio, depth and motion), increasingly capable Industrial AI models and physical systems such as robots and machines. Unlike traditional automation, physical AI enables systems to interpret unstructured environments, adapt to new situations and make autonomous decisions rather than following rigid, preprogrammed instructions.
While physical AI represents a solution to the problem of inflexible automation, achieving an AI model that can process complex multi-sensory data and reach reliable conclusions about actions to take is not an easy task. Collecting and managing the data required to train a physical AI model, along with an entire, real-world training environment would be impractical, if not impossible, without the comprehensive digital twin.
The Value of Physical AI
When it comes to the factory floor, one of a human’s greatest assets is an intuitive understanding of the world, which gives them the versatility to handle numerous odd and small tasks with minimal instruction. This includes tasks such as picking a screw from a bin or retrieving a box from a warehouse without running into anything. While trivial for a human, these tasks are surprisingly complex for an automation system.
Traditional automation relies on set paths and distances, requiring precise layouts of warehouses, strict frames to hold parts like screws and other rigid examples. While basic AI automation relaxes these restraints somewhat, allowing robotics to pick parts from a bin or a general area of a warehouse, these AI models fail to adapt to handling complex situations outside the range of their trained task.
This is why physical AI models are critical. They are trained to understand the physical world in general, not just a single task. Instead of being programmed for every step of a task or exhaustively trained on a new model, a physical AI model could simply be told what to do, even in natural language, and be left to complete its task. Much like a human would, it would be able to handle unexpected changes in the environment, adapt to emergencies and easily comply with new instructions, providing an unparalleled level of flexibility in automation.
To achieve this level of functionality, three areas are critical. The first two are the ability to sense and the ability to take action. The robot or machine needs to have depth perception sensing, audio sensing and the ability to move and navigate. This type of depth sensing capability is usually a result of cameras on the machine. And the object, such as a robot arm, needs to be able to touch, feel, etc., to perform a task. Both the sensing capability of cameras and the ability to execute actions has been rapidly improving over the last few years.
The final critical area for physical AI is a “brain” to oversee and operate all of this. This is the difference between traditional automation and robotics. With AI, physical objects can determine the course of action that they are going to take based on their reasoning capability, which can perform real-time inferencing.
The Digital Twin Accelerates Physical AI Training
Before physical AI can be applied to the factory floor, however, it must first be trained.
Training a physical AI model requires an environment where the model can learn the rules of the real world, from how sensors report data, such as how boxes fall when stacked improperly, to the very laws of physics. This kind of training is costly to do with physical systems. It requires dedicated robotics equipment, taking valuable capital equipment off the manufacturing floor and devising thousands or even hundreds of thousands of scenarios that involve the robot performing actions on actual physical objects.
Fortunately, a digital twin enables much faster, more comprehensive AI model training. A digital twin is a digital representation of a physical asset or process, and can be used to design, simulate and optimize products, machines, production and entire plants in the digital world before taking action in the real world. Using real-world data, the digital twin can accurately reproduce everything from factories and machines to humans and robotics. This provides an ideal environment to train an AI model.
For example, one can generate thousands of scenarios of picking mixed fasteners from a bin with varying part-pose distributions, clutter, reflectivity, lighting and camera noise, then train and validate the perception-and-grasp policy against those distributions. This approach allows the AI model to safely learn real physics without the need to produce large quantities of expensive real-world data, reducing the time, costs and resources required to train and validate the model.
Synthetic Data Provides Cost-Effective, Efficient Training
Synthetic data is one of the greatest benefits of digitalizing learning processes. It is any type of data generated using a computer instead of real data. Synthetic image data is particularly valuable in the manufacturing space where most tasks rely on accurate visual sensing capabilities.
Using synthetic data, thousands or even millions of unique sets of training data can be generated based on real world possibilities to quickly and automatically train a physical AI model. The environment’s properties can be randomized: from number of objects and their locations to material properties, camera properties, surrounding environment and more. Camera pixel resolution, noise, ambient lighting effects, etc., can also be added to the synthetic image generation to mimic actual behavior.
By training a machine learning model with such a randomized dataset, the model will learn how to ignore the properties that are randomized and focus on the ones that are not (such as the part geometry). As a result, the trained model can generalize to various environments and domains, including the real expected environment.
Synthetic “pictures” are only one type of synthetic data that can be produced using the digital twin. As a true-to-life replica of the real world, the digital twin can provide a virtual sandbox of the surrounding environment for the physical AI model to safely learn about the real world, including input from various types of sensors and interacting with that environment using manipulators and locomotion systems in which the robot will be operating.
Opportunities for Physical AI in Automation
As demands for automation systems with flexible, human-like decision-making abilities increase, the challenges of training the physical AI models required to make them a reality grow as well. But, with the continuing digitalization of the manufacturing industry and increasing adoption of the digital twin, a path to address these challenges exists.
In fact, physical AI is already being applied across several automation domains, monitoring machining conditions such as vibration and tool wear in machine tool operations, interpreting sensor data across complex production processes in process automation and detecting surface defects, dimensional deviations and assembly errors in real time in quality inspection. The number of applications will continue to increase as the push for smarter automation goes beyond building a single robot to perform a pre-determined task.
To keep pace with this demand requires systems that can think and adapt on the fly to provide flexibility in production and allow for greater customization and optimization. Future systems will be required to adapt their production processes from one day to the next without requiring a new training model for every single variation of the production line.
As physical AI continues to advance, it will increasingly be able to tackle tasks previously limited to humans, especially those requiring rapid adaptability in operations that traditional automation and AI systems could simply not handle. While training physical AI models may seem like a daunting task, by intelligently leveraging the digital twin, the process can be greatly simplified, enabling companies to reap the benefits of flexible, adaptive automation that will fundamentally transform their manufacturing capabilities.
This article was written by Rahul Garg, Global Vice President of Industrial Machinery, Siemens Digital Industries Software (Plano, TX). For more information, visit here .
Transcript
00:00:00 Here in the electronics factory in Alangan, we went from basically zero automation to more than 100 lightweight robots in use today. But they are all static. The idea of humanoid robots is that they can actually move around and are universal enough so they can do more than just one task. We have decided to partner with the
00:00:33 startup humanoid because their team brings strong industrial experience from leading companies and since their founding they have made impressive technological progress. They worked in a very structured way and they delivered a pragmatic solution for our manufacturing PC. Okay, so this is an inbound logistics operation. It happens across a lots of
00:00:54 different manufacturing facilities. It's effectively moving a tote from a stack and putting it onto a conveyor to be brought into the facility. We use a humanoid robot in that special PC because we want to see how humanoid robots behave especially in task where you have to for example bend over to take a box and perform humanlike activities.
00:01:18 It's a very arduous repetitive task that drives, you know, occupational health issues because you're lifting heavy loads uh repetitively. The reason we're here is to transition the robot from the lab to the line. In the next 12 months, we will move from a single robot operation within a proof of concept deployment such as this one to a multi-root deployment as part of a pilot
00:01:43 deployment in a live manufacturing facility doing real work for our customers. Every hour we get to operate in the customer's facility enables us to iterate faster and ensure we're staying ahead of the market. The role of the human on the shop floor will change because the robots will be able to do the redundant and unergonomic
00:02:24 tasks giving the human employees more room to pursue creative tasks or more value adding tasks.
>> The seaman's team were incredibly fast, very responsive and on day one we were able to uncreate the robots and we were picking boxes and totes almost immediately. It's been brilliant to work in the facility with them live in a production
00:02:44 environment uh and get to incorporate the machine into such a production facility.
>> We're pushing the boundaries here of what we're building and how we're deploying it. Um and we're excited to see where that that will go to
>> and that's down to the uh the great partnership that we've been able to build uh with the Seaman's team here.
