If manufacturing conditions remained fixed, this gap could be addressed by building a sufficiently large library of validated recipes. In practice, manufacturing continuously evolves. New alloys, composites, coatings, adhesives, and resin systems enter production. Suppliers change, material batches vary, tools wear, quality requirements tighten, and manufacturers introduce new geometries that alter how force, heat, and material flow through a part. Manufacturing data is therefore inherently conditional. A process recipe is valid within an operating envelope defined by the material, geometry, tool condition, equipment configuration, environment, and required outcome. When those conditions change, historical data may remain informative, but it is rarely sufficient on its own. We have discussed this in detail in GrayMatter Robotics’ article, Chasing Value, Not Volume.
Between a customer’s production requirement and the motion executed by a robot, a critical layer of manufacturing knowledge is what we call the process recipe. The customer specifies the end result e.g., remove a coating, reach a target roughness, eliminate a defect, or deposit a uniform layer of material. The process recipe determines the tool-to-part interaction needed to achieve it. Only then can the robotic system determine the tasks, motions, and control behaviors required to realize that interaction across the geometry of the part. Today, this translation is performed primarily by experienced process engineers. Their expertise is not simply a memory of successful parameter values. It is an internal model of the process: an understanding of how variables interact, which failure modes are plausible, where operating boundaries may lie, and what should be tested next when an experiment produces an unexpected result. Engineers formulate hypotheses, conduct trials, inspect outcomes, interpret deviations, and decide which experiment will provide the most useful next piece of information. As deployments expand across new materials, parts, and applications, this expert-guided experimentation becomes the constraint. The scarce resource is no longer the ability to generate robot motions. It is the ability to efficiently discover the process knowledge that the trajectory must execute.
Curie is GrayMatter Robotics’ AI agent for process engineering. It turns a robotic cell into a controlled experimentation environment where process parameters can be varied, outcomes can be measured, and every result can inform the next experiment. Curie is not a single model that predicts an optimal parameter set. It is an integrated system for reasoning about process physics, conducting experiments, interpreting observations, and converging on a production-ready recipe. It combines four tightly coupled capabilities: a Process World Model, an autonomous experimentation environment, hypothesis generation, and informative experiment selection. The Process World Model represents how process inputs and tool-to-part interactions influence manufacturing outcomes. Depending on the application, it may capture relationships among force, velocity, tool condition, material properties, temperature, and standoff distance, as well as outcomes such as material removal, surface quality, defect formation, coating uniformity, and cycle time. The model is not assumed to be complete. It is a structured approximation that can be updated as new evidence becomes available.
The autonomous lab allows Curie to test that understanding in the physical world. A robotic cell equipped with appropriate sensors conducts controlled trials, while capturing the commanded parameters, tool-to-part interaction, and measured outcome. Each trial becomes a structured observation that can be compared with the Process World Model’s prediction. When an observation differs from the prediction, Curie can generate physically grounded hypotheses about the discrepancy. E.g., A new material may exhibit spatially varying compliance, or a previously secondary variable may become significant under the new operating conditions. Curie then selects the next experiment. The goal is not merely to test the parameter combination currently predicted to perform best. Curie also considers the information value of the experiment, whether it can distinguish between competing hypotheses, reduce uncertainty in a critical region, expose an operating boundary, or improve confidence in the emerging recipe. Curie allows process engineers to explore broader process spaces while concentrating their attention on decisions that require expert judgment.
No simulation fully captures the physics of a new manufacturing process. Material variability, tool wear, sensor noise, environmental conditions, and previously unseen failure modes ultimately reveal themselves through physical interaction. At the same time, exploring the entire process space in the real world would be slow, expensive, material-intensive, and potentially unsafe. Curie addresses this through a Real-to-Sim-to-Real learning loop. When Curie encounters a new material, tool, or production requirement, it begins with a limited set of carefully monitored physical experiments. These trials characterize the material response, establish a preliminary operating envelope, and reveal where the Process World Model diverges from observed behavior. This grounds the model in the actual process rather than assuming that prior data will transfer without modification.
Curie then moves into simulation. The real-world observations condition and refine the Process World Model, allowing the system to explore a broader range of conditions than would be practical through physical trials alone. Simulation can screen competing hypotheses, study interactions among variables, and investigate conditions that may be destructive, unsafe, or too material-intensive to test initially on physical coupons. The objective is not to declare a final recipe in simulation. It is to determine what remains unknown and identify which physical experiment would reduce that uncertainty most effectively. Curie then returns to the real world with targeted experiments selected for their information value. A trial may discriminate between competing explanations, validate a predicted boundary, test a promising region of the recipe space, or expose a mismatch between simulated and physical behavior. Its outcome is fed back into the Process World Model, and the loop repeats. Real-world grounding → Simulated exploration → Informative physical validation → Model refinement. Through this cycle, Curie narrows the recipe space until the process satisfies its quality, safety, throughput, and robustness requirements. The goal is not to eliminate physical experimentation, but to extract more knowledge from every experiment and converge using a limited number of trials. Curie does not simply generate a static process recipe. The recipe defines the policy governing how a specific tool should interact with a given material and geometry to achieve the required process outcome. Curie uses this process policy to post-train GrayMatter Robotics’ task and skill models for the customer’s application, adapting them to the relevant part geometry, tooling, process constraints, and operating environment.
Curie does not replace the process engineer; it expands the scale at which that engineer can operate. Human experts continue to define objectives, safety constraints, operating boundaries, and production acceptance criteria, while Curie executes controlled trials, analyzes outcomes, and identifies the experiments most likely to advance understanding. This enables one engineer to supervise broader experimental campaigns, evaluate more materials, and qualify more applications using the same expert time and infrastructure. By turning process-knowledge generation into an AI-accelerated capability, Curie shortens the path from a new material or customer requirement to a production-ready robotic process.
This case study focuses on a carbon-steel elbow flange (Fig. 1b), a common component in marine applications. The production requirement was to remove all rust, achieve 100 percent coverage, and satisfy the cleanliness specification. Before Curie, qualifying a comparable process relied largely on Design of Experiments conducted by an experienced process engineer. Each trial consumed coupons, abrasive media, machine time, and engineering hours. Trials proceeded sequentially, with inspection and interpretation between rounds, while the engineer determined what to test next. A qualification of this type typically required two weeks and 38 physical trials to produce a recipe acceptable for customer production.
We ran the same recipe generation campaign with our Process Engineering Agent: Curie. Curie queried the Process World Model for the regions of the blasting parameter space where predictive uncertainty was highest for this material, then generated an initial experiment set on 8in x 8in carbon steel coupons made from the same material as the production part (Refer Fig. 1a). Each trial recorded the commanded parameters, the tool-to-part interaction, and the measured outcome (material removal, anchor profile, coverage), giving the model a structured observation to compare against its prediction.
Figure 1a: Test coupon and elbow flange blasting results
Note: The test coupon is 2ft x 2ft and for trials it is split into 9 sections of 8in x 8in each.
Figure 1b: Before vs. after blasting comparison
With these observations, Curie delivered a qualified, production-ready recipe in 2 days, compared to approximately two weeks for the same qualification run manually. The campaign required 12 physical trials against a baseline of roughly 38, with the remaining exploration absorbed by the Process World Model in simulation. Refer to Fig. 1b for the performance of the learned recipe.
| Manual Baseline | Curie Campaign | |
|---|---|---|
| Qualification Time | 2 weeks | 2 days |
| Physical Trials | 38 coupons | 12 coupons |
Three mechanisms drive these gains. First, informative experiment selection means each physical trial is chosen to reduce uncertainty or discriminate between hypotheses, so fewer trials carry more knowledge. Second, simulated exploration absorbs the broad parameter sweeps that previously consumed coupons, media, and floor time. Third, hypothesis-driven iteration replaces undirected trial and error.
