Dr. Satyandra K. Gupta: Co-Founder, Chief Scientist, GrayMatter Robotics; Smith Int’l Professor & Dir. CAM USC; Fellow ASME, IEEE, SME, SMA.
Most of the AI that we experience in our daily lives is digital AI. It produces digital artifacts, decision recommendations or predictions that will either be used by a human or some other digital agent. Examples include generating a cover letter for a job application using ChatGPT, recommendations for watching a movie on Netflix, creating a painting using Dall-E and detecting a tumor in a medical image.
But a different kind of AI—called embodied AI—is being developed to manage the behavior of physical systems. It’s tasked with producing a sequence of actions that the physical system executes to achieve the goal. For example, a robotic cell can be tasked with sanding the top surface of a part placed in the cell with the desired surface finish. The embodied AI monitors the cell state using sensors and generates instructions for the robot to perform the task.
Digital AI and embodied AI share some similarities and utilize many underlying techniques. However, understanding the differences between these two types of AI is critical to successfully adapting digital AI approaches for use in the context of embodied AI applications.
The risk profile of embodied AI applications is often fundamentally different from that of digital AI applications. If the accuracy of a digital AI tool is 99%, it can tremendously boost human productivity in many applications. For example, if you use generative AI—one type of digital AI—to generate a 1,000-word cover letter that only requires you to manually edit 10 of those words, you’ll save a lot of time compared to writing that letter from scratch. And for the case of a recommendation engine, you wouldn’t mind if it gave you a poor suggestion for a movie once every couple of months. Comparatively, accuracy requirements for the embodied AI system are often very different due to risk considerations. For example, if a robot has a success rate of 99% on processing steps, and it works on a part that requires 200 steps, then every part made by the robot will contain two errors. As a result, the part would get scrapped or would require repairs. In most manufacturing applications, this isn’t a viable technology.
Risk consists of two aspects: the probability of making an error and the consequence of making errors. When the consequence of making an error isn’t significant, a much higher probability of error can be tolerated. That’s why an error probability of 1% will be acceptable in many digital AI applications. Conversely, many embodied AI applications demand error probabilities to be better than one in a million. Reducing error probability using a purely data-driven approach requires enormous amounts of data. In most cases, the need for data grows exponentially. Unfortunately, acquiring data from physical systems is expensive. Therefore, a different approach needs to be followed when dealing with embodied AI applications.
To address the requirements outlined above, embodied AI for manufacturing applications needs to have the following characteristics.
• Trainable With Limited Data: The embodied AI should be designed to be trained with limited data generated by physical experiments.
• Composable From Pre-Trained Modular Components: Physical systems can have a wide variety of configurations to support their intended requirements. For example, manufacturing robotic cells can come in many different configurations based on the process being performed (e.g., sanding or blasting). Different cells may include robots with varying capabilities (e.g., mobile platform-mounted robot or gantry-mounted robot), types of sensors (e.g., depth camera or thermal camera) and tools (e.g., orbital sanders or blasting nozzle). Therefore, developing universal embodied AI that works for all manufacturing applications out of the box isn’t likely to perform well. The AI for the system needs to be quickly synthesized from the modular components to match the sensing and actuation capabilities of the system and the work environment.
• Adaptable Based On New Data Or Context: As new data becomes available by deploying the system, it should be possible to improve the performance of the AI by using that data. AI should be able to adapt to new environments or tasks autonomously with minimal human supervision.
• Upgradable With Minimal Effort: The performance of a physical system may change over time because of wear or updates to physical components. This may require you to refine the AI to ensure that it can keep up with system evolution. Therefore, the embodied AI system needs to be designed to ensure that it can be updated with minimal disruption to the system operation.
• Risk-Informed Action Recommendations: The system should be able to estimate its confidence in recommended actions. Low confidence should prompt the system to perform risk analysis to analyze the consequences of failure. If the risk is too high, it should seek help from humans.
• Explainable: If the system-recommended actions don’t match user expectations, the system should be able to explain the rationale used to select actions.
• Distributed Architecture To Enable Computation Partitioning Between The Edge And The Cloud: It’s impossible to perform all AI computation in the cloud in the context of embodied AI. The system should be designed to ensure that computation that’s sensitive to network latency can be performed on the edge.
In digital AI, we see great success with large end-to-end learning models (e.g., LLMs). These models thrive on vast amounts of data. However, they don’t have many of the characteristics of embodied AI mentioned above.
Embodied AI should be viewed as a complex system that involves interactions among multiple AI components. Having the right system architecture in the embodied AI is the key to success in manufacturing applications. This allows you to exploit the recent advances in AI and meet the demanding requirements of manufacturing applications. Therefore, a modern systems engineering approach needs to be used to design the embodied AI for manufacturing applications.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
