Addressing Manufacturing Challenges with Automation

Q: how does physical ai work?

Unlike digital AI — used in large-language models like ChatGPT and Gemini which operate solely in the digital realm for data analysis and predictive tasks — physical AI interacts with the physical world. Sami Haddadin, a robotics and AI researcher, describes embodied (physical) AI in a Columbia University article: “The key difference is that embodied AI learns through experience and interaction, much like humans. It builds models of the world through sensory feedback and real-world interaction rather than just analyzing statistical data.”

Every Friday at 3 p.m., paint‑shop supervisor Ryan Steffen walked the line at Life Line Emergency Vehicles, hunting for volunteers to sand ambulance panels. Dusty, wrist‑numbing work — no one rushed to raise a hand. Then, one Thursday, a yellow‑and‑black robot rolled onto the floor. By the next Friday, Ryan wasn’t searching for volunteers; he was checking finished panels.

Award-winning (RBR50) Scan&Sand solution at Life Line Emergency Vehicles, sanding sections of the vehicle which previously were very hard for operators to process. Now with the solution, operators are able to make the robot take on the brunt of the work, freeing them up for more valuable tasks around the shop.

Automating processes in high-mix shops has historically been a challenge, and surface finishing with complicated geometries even more. But GrayMatter Robotics (GMR) offers turnkey automation solutions focused on finishing operations and quality inspection that are tailored for these high-mix, high-variability production shops, including those in aerospace and defense and specialty vehicle production. Requiring zero capital investment upfront, these physical AI-powered systems can be fully operational after one to two days of in-house training, providing a significant edge over competitors.

Founding GrayMatter Robotics

Founders SK Gupta, Brual Shah and Ariyan Kabir.

GrayMatter Robotics was founded by Santyandra K. (SK) Gupta, Ariyan Kabir and Brual Shah, who met as researchers at the University of Maryland’s engineering school. Drawing on their AI and robotics expertise and real-world manufacturing experience, they recognized the potential for physical AI-powered robotics to streamline manufacturing workflows. They envisioned a user-friendly, turnkey system requiring no programming, and operating on local computers without internet or cloud access. After collecting highly valuable data by conducting many tests on complex parts, they developed the technologies now powering GrayMatter Robotics’ commercial solutions.

Why manufacturing AI isn't like ChatGPT: The critical difference

GrayMatter Robotics leverages physical AI to deliver robotic solutions that enhance productivity and quality in finishing and inspection applications. Unlike digital AI — used in large-language models like ChatGPT and Gemini which operate solely in the digital realm for data analysis and predictive tasks — physical AI interacts with the physical world. Sami Haddadin, a robotics and AI researcher, describes embodied (physical) AI in a Columbia University article: “The key difference is that embodied AI learns through experience and interaction, much like humans. It builds models of the world through sensory feedback and real-world interaction rather than just analyzing statistical data.”

And according to GrayMatter Robotics experts the risk profile of physical AI applications is often fundamentally different from that of digital AI applications. A digital AI tool with an accuracy of 99% is useful, as it can tremendously boost human productivity in many applications. For example, if a user creates a document using digital AI that only requires them to make small changes, they have saved a lot of time while incurring very little risk.

However, physical AI in manufacturing has much stricter 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 is not a viable outcome.

With physical AI, the error probabilities are of a different magnitude, such as one error in a million. Enormous amounts of data are required to calculate risk, and the need for continual data inputs means the data being processed will grow exponentially, taxing the computer networks and growing expensive. To avoid this, a different approach needs to be taken when implementing physical AI applications.

How GMR’s physical AI works

GrayMatter Robotics pioneers physical AI in manufacturing, using it to tackle tasks like sanding, grinding, blasting, painting, polishing and inspection. Its modular systems autonomously scan parts and begin operations with minimal operator input. These systems are also affordable and quick to install, making them ideal for shops new to automation.

“Our AI-driven robot uses vision via cameras and sensors to build models of parts placed in front of it,” said Brual Shah, CTO and cofounder of GrayMatter Robotics, during a system demonstration. “Shopfloor operators handling high-mix operations can place parts on any transport cart in any orientation. As long as the cameras can see and digitally map all surfaces, the robot generates the necessary motions to scan and finish the part. Operators press one button, and the robot begins scanning and sanding.”

A smart robotic solution that autonomously sands large emergency vehicles, seamlessly fitting into Life Line’s limited shop floor. It tackles repetitive, labor-intensive sanding, freeing skilled operators from uncomfortable, fatigue-inducing tasks.

This system was developed through experiments and trials using contact kinematics, studying movement, velocity and pressure, changes in materials to determine optimal tool-to-part interactions. The computer generates operational instructions on an edge computer located at the machine, eliminating network latency and risks. GMR’s black-box solution is self-contained, requiring no cloud connectivity — a critical feature for sensitive industries like aerospace and defense.

GMR’s physical AI performs sequential, manual actions to achieve goals. For example, metal or composite components are scanned using integrated cameras and sensors that capture complex geometries. After entering a few parameters, the system’s cameras, laser scanners and sensors, mounted on robotic arms, use laser beams to intersect with the part’s 3D surface, generating a program on the fly. The system identifies safe contact areas on each part and creates instructions that help avoid potential damage to the part or the robot.

Once scanned, the processor creates instructions, enabling immediate operation. And the system can be trained to replace its own consumable tools and parts during operation, allowing it to run unstaffed for hours at a time.

Once operations are completed, the same system can be used for quality control with GMR’s Scan&Inspect™ module, which executes dimensional inspection on large and complex parts for metrology and defect detection or for online process monitoring purposes.

Physical AI requirements for manufacturing

To better explain what physical AI requires, the experts at GrayMatter Robotics defined the following for manufacturing applications:

Trainable with limited data: Users must be able to train the model and thus robots with limited data generated by physical experiments.
Utilize pretrained modular components: Due to the wide variety of configurations in physical systems, a universal physical AI that works for all manufacturing applications in a facility is not likely to perform as needed. The AI for the system needs to be quickly synthesized from modular components to match the sensing and actuation capabilities of the system and the work environment.
Adaptable: New data will be collected after deploying a system, so that system should be adaptable to incorporate this new information autonomously with minimal human support.
Upgradable: Physical components can wear out or may need to be upgraded for a multitude of reasons. The system should be able to be modified with minimal disruption.
Risk-informed action recommendations: The system should be able to estimate its confidence in recommended actions. If the risk is too high, then it should seek help from humans.
Explainable: If the system-recommended actions do not match user expectations, then the system should be able to explain the rationale used to select actions.
Local computational ability: Because many applications of physical AI need to occur in a local system, systems should be designed with that in mind.

Having the right system to interact with multiple AI components and the changing needs of manufacturers must be considered when designing a physical AI system or workflow for manufacturing, and that’s where GrayMatter Robotics is succeeding.

From skepticism to success in two days

Life Line Emergency Vehicles in Sumner, Iowa, has been using GrayMatters Robotics Scan&Sand™ in the production of its customizable ambulances. The company operates two shifts daily, five days a week, and each ambulance requires three sanding operations in the paint department alone, a repetitive task that was a constant source of staffing headaches and a potential bottleneck in their workflow.

Finding and retaining skilled technicians willing to sand for hours was a persistent challenge. Ryan Steffen, Life Line’s paint supervisor, says management explored various automation options, but many required significant customer-side programming and high capital investments, and would still fall short of Life Line’s needs. GrayMatter Robotics’ Scan&Sand™ system changed that.

Life Line has processed hundreds of vehicles using the Scan&Sand™ solution on three substrates: raw aluminum, primed aluminum and final paint finishes. The system reduced sanding time by more than 30% and lowered consumable use compared to manual processes. Steffen says its sandpaper has also lasted longer than expected, prompting GMR to analyze usage data and adjust the robot’s operation to reduce changeouts, saving time and costs.

“I’d tell anyone considering automation to embrace increased production. It doesn’t take away jobs; it handles monotonous, taxing tasks.”

From the initial conversations with GMR, Steffen says Life Line was impressed with the company’s sales and implementation process, customer service and responsiveness. The decision came down to seeing it in action.

Life Line sent one of its production parts to GMR's facility. They watched as the system scanned the unique geometry and, within minutes, began sanding — all without a single line of custom code. That live demonstration, seeing their own part finished flawlessly, removed any doubt and was the proof they needed to move ahead with GrayMatter Robotics.

Steffen praises GMR’s exceptional customer service and regular updates to both software and hardware based on information gained across their installed base. One example he shared was an update to their camera brackets to stabilize them.

Steffen acknowledges his initial hesitation about adopting robotic automation but says benefits were immediate. “It can be scary to incorporate new technology, but keep an open mind,” he urges others. “It’s not about replacing people or jobs but making them easier to get done. Few people want to sand, and GrayMatter Robotics’ Scan&Sand™ solution eliminates hard labor and repetitive motion, reducing physical stress on our team. I’d tell anyone considering automation to embrace increased production. It doesn’t take away jobs; it handles monotonous, taxing tasks.”

Flexible configurations for high-mix production

Each shop’s unique setup and product mix require tailored solutions. GMR offers multiple base configurations with customization options and rail lengths including:

a single-sided loading system with one rail or pedestal for sanding, buffing or polishing with a large work envelope available in 4-, 8- or 12-meter rails;
a dual-arm system with two rails and two robots for high-efficiency, center-loaded operations on larger projects;
and a mobile arm and rail system for large horizontal surfaces, ideal for low-force tasks like sanding and spraying.

The world’s first dual-armed autonomous sanding system, simultaneously prepping two large marine parts.

One GrayMatter Robotics cell can perform multiple operations — sanding, grinding, blasting, polishing, painting and inspection — reducing the need for multiple cells and saving space and capital investments while freeing employees to perform more high-value tasks. The cell configuration is the same from one system to another, only requiring different tooling and operation-specific ancillary equipment like dust collectors. The company can help users determine the best system for their shop and help customize the setup needed to get operational as quickly as possible.

GrayMatter Robotics provides ongoing training for operators and super users, ensuring your team is equipped for success as your workforce evolves. Our training is flexible, including remote sessions to accommodate your team’s schedule and in-person training to ensure a comprehensive understanding of our robotic solutions.

Looking ahead

GrayMatter Robotics continues to innovate, expanding its capabilities to solve new manufacturing challenges across the U.S. and globally, aiming to make automation accessible and affordable for all.

For more information on today’s offerings or to learn about what’s coming from GrayMatter Robotics, check out the company’s website.