Physics-Informed AI for Powering Smart Robotic Cells in Manufacturing Applications
by Satyandra K. Gupta, Co-Founder & Chief Scientist at GrayMatter Robotics
Challenges in Using Data-Driven AI in Manufacturing Applications
Over the last few years, data-driven Artificial Intelligence (AI) has delivered very impressive results in wide-ranging applications such as recommendation engines, playing games, face recognition, text translation, text synthesis, and fraud detection. This type of AI uses a vast amount of data to train the system. In many applications, a large amount of data is readily available (e.g., images on Instagram, a large amount of text on Wikipedia pages) or can be easily generated (e.g., a computer playing a game against itself). In contrast, collecting high-quality data in many manufacturing applications takes significant time and incurs large costs. Therefore, unfortunately, a purely data-driven AI approach is not a scalable model in many smart manufacturing applications. We need a different type of AI.
Manufacturing has a lot of known models and useful process knowledge. Rediscovering these models and knowledge using a purely data-driven approach does not make sense. However, all known models make simplifying assumptions to reduce complexity and therefore are approximate in nature. We need AI that exploits the known models and uses a data-driven approach to augment the known models and existing knowledge based on experimental data to fill the missing gaps. This type of approach is called Physics-informed AI. It enforces known physics-based process models (or knowledge) as a constraint in the AI system to ensure that it does not learn anything that contradicts existing models/knowledge. For example, the system can enforce a constraint that increasing pressure on the sanding tool will increase the deflection of the part being sanded. We don’t need to conduct a large number of tests to learn this already-known fact. If the measured data contradicts this constraint, then it is highly likely either the sensor is malfunctioning, or the part/tool is not clamped properly.
On one hand, the physics-informed AI approach restricts the solution space and, therefore, makes the problem much more tractable from the data requirement point of view. Let us consider the problem of predicting process output based on the input. If the output is expected to increase with an increase in the input, then the underlying model space is limited and it can be trained by a smaller amount of data because we don’t need to consider arbitrarily complex models. On the other hand, using physics-informed AI requires more complex representations and associated methods to handle constraints to produce acceptable computational performance. We cannot use a simple neural network and train it with observed input and output data. If we did not explicitly enforce process model constraints, then there is no guarantee that the learned model would preserve the process constraint if the output used during training is noisy.
Physics-Informed AI Use Cases in Smart Robotic Cells
Here are a few representative use cases for using Physics-informed AI for realizing smart robotics cells in high-mix manufacturing applications.
- Defect detection is an essential component of deploying smart robotic cells in manufacturing applications. Machine learning has emerged as a powerful technique for analyzing and classifying images. However, collecting a large number of images of physical defects needed to train a machine learning system is not possible. An alternative is to develop a pipeline for generating photo-realistic images to overcome this challenge. In this case, a physics-based process model can be utilized to generate realistic defects in virtual models. These virtual models can be used to generate photo-realistic images. Recent work in this area has demonstrated that a training process that utilizes a combination of photo-realistic synthetic images and real images of defects works well in practice.
- To efficiently and accurately build part models for autonomous finishing (e.g., sanding and polishing), a machine learning-based approach is being utilized for predicting sensor performance (e.g., measurement errors) as a function of the operating conditions. This approach needs to use known models of sensor performance during the training process. For example, if the sensor is expected to produce a higher amount of error when imaging from a larger distance, then this information is used during the training phase to ensure that the image is acquired from the right distance. This approach enables the acquisition of good-quality data for improved decision-making during autonomous robotic finishing.
- Robots often require body-mounted hoses and cables for the cell to function. These appendages may restrict the motion of the robot. Applying overly conservative restrictions makes the robotic cell inefficient. On the other hand, not applying appropriate limits may result in damage to cables/hoses and force the cell to shut down. A learning framework can enable users to estimate robot motion limits when flexible components are attached to the robot. This approach can use a physics-based model of cables and hoses bending, twisting, and stretching to accelerate the learning process.
- A smart robotic cell should be capable of building process models for new materials by autonomously conducting experiments. While the exact quantitative relationship between the input process parameters and process performance may not be known, often qualitative relationships between many variables are known (i.e., increasing torch velocity decreases weld height). We can utilize loss functions during the training phase that penalize deviations from known process constraints. This approach can enforce known knowledge and accelerate the model-building process to enable autonomous processing.
- Smart robotics cells need to use advanced prognostics and health management to ensure a higher level of reliability. We need to utilize known causal relationships during the inferencing process to ensure that we can accurately predict the system state based on the process monitoring data. For example, we can utilize the force and vision data to hypothesize the cause of accelerated tool wear in robotic sanding.
The above examples illustrate how physics-informed AI technologies are enabling the realization of smart robotic cells to function autonomously and deliver consistent quality. Smart robotic cells can also monitor their own health and reduce the risk of failure by proactively taking corrective actions. These capabilities are expected to revolutionize the field of manufacturing by bringing automation to high-mix manufacturers.
Keep up with us:
Satyandra K. Gupta
Co-Founder & Chief Scientist at GrayMatter Robotics
Dr. Satyandra K. Gupta is the Co-Founder and Chief Scientist at GrayMatter Robotics. He holds the Smith International Professorship in the Viterbi School of Engineering at the University of Southern California and is the founding director of the USC Viterbi Center for Advanced Manufacturing. Formerly a program director for the National Robotics Initiative at the National Science Foundation, Dr. Gupta's research focuses on physics-informed artificial intelligence, computational decision-making foundations, and human-centered robotics. With over 400 technical articles, 180 invited talks, and recognition as a fellow in prominent engineering societies, he contributes significantly to the field. Dr. Gupta has also garnered media attention, featured in outlets like Economist, Forbes, Huffington Post, LA Times, and Smithsonian Magazine.