Is AI Ready to Tackle High-Mix Grinding? Interview with GrayMatter Robotics Co-Founder

An AI-powered robotic grinding system performs a weld-blending operation at GrayMatter Robotics’ California facility. The system uses force control and 3D scanning to adapt in real time to part geometry and material variation. All photos provided by GrayMatter Robotics.

Dr. Satyandra K. Gupta has focused his career on pushing the boundaries of robotic intelligence. In addition to co-founding GrayMatter Robotics (GMR) in 2020, Gupta is the Smith International Professor of Mechanical Engineering  and Computer Science at the Viterbi School of Engineering, USC — and the founding director of its Center for Advanced Manufacturing. In each of these roles, Gupta has devoted years of research into to developing AI systems that can perceive, plan and act precisely in the physical world. In this Q&A, Gupta discusses how GrayMatter Robotics has translated that vision into an AI-driven robotic grinding system capable of handling high-mix production environments.

This interview has been lightly edited for length and clarity.

Modern Machine Shop (MMS): What inspired you to co-found GrayMatter Robotics? What specific problem in manufacturing you were trying to solve?

Satyandra K. Gupta: I’ve been working in robotic automation for a long time. At my lab at the University of Southern California, we were looking into how robots can be useful in sanding, assembly, machine tending, inspection and composite layup. Companies kept coming to us and asking for solutions specifically in surface finishing — sanding, polishing, grinding — because there’s an immense labor shortage and the work is ergonomically unsafe. People don’t want to do this work, and quality has become a major problem.

So we explored different ways of delivering solutions. But the high-mix applications we were targeting required AI-powered robots, and that wasn’t the approach system integrators were going to use. That’s why we launched the company. We started with sanding — a more forgiving application — and then added polishing, sandblasting, and eventually grinding.

MMS: Let’s talk about the kind of AI GrayMatter is developing. Can you describe the core principles behind GMR-AI and how it enables autonomous robotic grinding?

Dr. Satyandra K. Gupta is the co-founder of GrayMatter Robotics and the Smith International Professor of Mechanical Engineering and Computer Science at the University of Southern California.

Gupta: The process begins by scanning a part, because in high-mix, you can’t rely on CAD models alone. There may be differences between CAD and what appears on the shop floor, especially with castings or forgings. So we scan the part using laser or 3D imaging technology and build our own part representation.

That’s the first role AI plays: building a model of the part under varied lighting and surface conditions. Next, the system has to understand where the grinding needs to take place. Sometimes this is marker-based, where the user specifies a target region. Again, AI helps interpret that input.

Then the robot programs itself. It plans the motion — how the grinding wheel will approach, make contact and perform the operation. That’s the third way AI is used. During operation, the robot monitors the process: force feedback, unexpected behavior. AI evaluates that in real time and can bring the robot to a safe state and alert the operator if something goes wrong.

And finally, for any new metal, the system builds a process model. It learns what RPM to use, how much force to apply and how much material is removed. So AI plays a role in five ways: scanning, identifying regions, motion planning, process monitoring and process modeling.

An operator resets a grinding cycle using GrayMatter Robotics’ Scan&Grind interface, designed for ease of use by shop floor personnel without robotics experience.

MMS: What is embodied AI, and why is it critical for grinding tasks compared to more conventional digital AI?

Gupta: Embodied AI, or physical AI, is different from digital AI. Digital AI produces digital output — text, images, video — for humans or computers to consume. Nothing physical happens. But embodied AI recommends actions that are executed in the physical world, often without human review.

So let’s say you use digital AI to write a cover letter. If four words out of a thousand are wrong, you edit them. No problem. But with embodied AI, if even one motion segment of a robot path is wrong and if the grinding wheel crashes into the part, it’s a disaster.

The risk and acceptance criteria for embodied AI are completely different. That’s why grinding, which requires high physical precision and control, depends on embodied AI to work safely and reliably.

A robotic arm equipped with a belt grinder performs a grinding operation on a curved metal surface. The system uses force control to maintain consistent contact pressure throughout the process.

MMS: GrayMatter also uses physics-informed AI. How does that help with adapting to different alloys and surface irregularities?

Gupta: Physics-informed AI is a type of embodied AI. It means the AI has to obey the laws of physics — for example, if you apply more force, you remove more material. Even if data gets corrupted, AI can’t just make up behaviors that violate physical law.

When grinding different materials, the robot conducts offline experiments. It teaches itself: “How do I grind aluminum? How do I grind steel?” It learns the tool/material interaction, guided by physical principles like RPM, pressure and heat generation. That’s for material variability.

For geometry, the scanning process handles variability in shape. You might present one casting, then another — the robot scans and understands what’s in front of it. That’s how it handles shape variability.

MMS: How dynamic is your AI-driven path planning during a grinding pass?

Gupta: Once a path is computed and execution starts, the system doesn’t replan the path in real time. But it does adapt the path perpendicular to the contact using a force sensor. The tool adjusts up or down to maintain correct contact force. After the pass is complete, the system can re-scan the surface and generate a new path if needed.

MMS: What types of grinding applications is GMR’s “Scan&Grind” targeting today, and what’s next?

Gupta: Today, we’re targeting light surface finishing, weld blending, and cutting operations like removing parting gates. Soon, we’ll offer heavier material removal capabilities as well.

MMS: Walk us through the scanning process and how you ensure accurate modeling of complex geometries.

Gupta: We use commercially available laser line scanners or 3D imaging systems. The robot performs one or more passes to scan the part. If a single scan isn’t enough, it performs additional scans until the coverage is sufficient.

These technologies give us 50 to 100 micron accuracy — good enough for light surface finishing and weld blending. If an application in the future requires tighter tolerances, we’d explore higher-precision scanning technologies.

A GrayMatter Robotics system uses a rotary tool to test the removal of gates from cast metal parts. The AI-driven robot adjusts toolpath and force in real time to ensure repeatable results across varying part geometries.

MMS: What about thermal distortion? How does Scan&Grind deal with heat during grinding?

Gupta: In some operations like gate removal, heat isn’t a major issue. But for weld blending or surface finishing, we pay close attention to heat generation during process development. We regulate the material removed per pass and avoid excessive force. It depends on the material, too — conductivity, heat diffusion. So the recipes are carefully tailored offline for each material.

MMS: How precisely can you control contact force?

Gupta: Very precisely. We use a high-resolution force sensor that reacts at kilohertz frequency. It adjusts the tool’s vertical position to maintain target force.

MMS: What about the “feel” aspect of grinding — how can a robot approximate human intuition?

Gupta: Our system measures force better than a human can. After each pass, it can re-scan to assess the result. Humans may have more sensing modalities like tactile or visual inspection, but we use dimensional measurement and force feedback. Surface roughness measurement is also possible, but not yet in-process.

MMS: Can Scan&Grind detect and compensate for tool wear?

Gupta: For grinding, we currently estimate tool life based on usage time. For sanding, we have vision-based detection and measurement, and we plan to bring that to grinding soon.

MMS: How does the system handle safety — both human and mechanical?

Gupta: For human safety, all grinding cells are fully enclosed with physical barriers and access control. For robot safety, the force controller ensures that if force builds beyond limits, the robot retracts automatically.

A GMR robot performs surface-finishing operations on a series of identical metal assemblies. The system is mounted on rails for sequential processing and consistent treatment across multiple faces, with no need for manual repositioning.

MMS: What does success look like for GrayMatter in the next five years — specifically in metalworking?

Gupta: We want to be the tool of choice for a wide range of metal industries. We’re expanding into cutting and light machining as well. Broad adoption and acceptance of our technology — that’s the goal.

MMS: Do you see a broader role for AI in subtractive manufacturing beyond grinding?

Gupta: Absolutely. Robots are becoming more rigid, so machining is a growing area. AI can play a huge role in quality control, especially with enhanced sensors — detecting defects upstream and downstream. Eventually, AI will even help design new cutting tools using generative models.

MMS: Last question. How do you respond to skeptics who say robotic grinding will never match the precision of, say, a double-disc grinder?

Gupta: First, we try to understand where the skepticism comes from. Often it’s about sensing — surface finish, in-situ feedback or dimensional accuracy. In many cases, the technology can match human performance, but the cost of integrating advanced sensing may be too high for certain applications.

There are three categories: tasks where robots perform competitively at a good price point, tasks that require better sensing than what we have today, and tasks where we could match performance, but not cost-effectively. With rational discussion, most people see that it depends on the use case — and that the technology is advancing quickly.

Request A Demo With GrayMatter Robotics

Interested in learning more? Schedule a 45-minute demo with GrayMatter Robotics.