World Models in Manufacturing
Physical AI has made substantial progress in enabling robots to perceive their environment, reason about object geometry, and generate increasingly capable motions. But manufacturing requires another layer of intelligence: predicting what a physical action will do to the part. In a manufacturing process, the relevant action is not simply a robot trajectory, but the physical interaction between tool and part created by parameters such as tool velocity, orientation, contact force or stiffness, dwell time, abrasive grit, standoff distance, tool condition, and other process- and material-specific variables. These actions determine manufacturing outcomes that differ across processes, e.g., surface roughness in sanding, coverage in abrasive blasting, coating thickness in spraying, or material removal in grinding. A robot that cannot predict how its actions influence these outcomes can execute motion, but it cannot reason directly about whether that motion will produce the desired manufacturing quality.
Why Motion Alone Is Not Enough in Manufacturing
Modern robotic systems are increasingly capable of understanding object geometry, planning trajectories, and adapting motion in complex environments. But in manufacturing, executing the right motion does not guarantee the right outcome. The same trajectory can produce very different results depending on factors such as force, velocity, tool orientation, material properties, tool condition, or dwell time. For example, a sanding robot may follow the correct path over a surface yet still remove too much material, leave an unacceptable surface roughness, or fail to achieve uniform coverage. What matters is not only where the robot moves, but how that interaction changes the part. A Process World Model captures this relationship by learning how the current process state, the chosen process action, and the surrounding physical context determine the resulting state of the part: (St, At , Ct) → St+1. where St is the current state, At is the candidate tool action, Ct represents the relevant physical and process context, and St+1 is the predicted state after that action is applied. This gives the robot a predictive model of the manufacturing process itself. Geometry and motion planning determine what the robot can do; the Process World Model helps predict what those actions will produce.
A World Model for Manufacturing Process
Most world models in robotics are designed to predict future observations: the next image, sensor reading, object configuration, or latent state of the environment. Manufacturing requires a different type of prediction. The system must estimate how a process action will change the manufactured part itself.
For a process such as sanding or blasting, the current state of the part and tool can be represented spatially as: St (x,y). Given a candidate process action At and the relevant physical context Ct, the Process World Model predicts the resulting state: S t+1 =f(St, At, Ct) . The predicted state can represent manufacturing-relevant quantities such as material removal, surface roughness, process coverage, coating thickness, or other quality-critical outcome fields, depending on the process being modeled. This distinction is important because many robot actions may be geometrically valid while producing very different manufacturing results. A motion planner can determine whether an action is executable; the Process World Model predicts whether that action is likely to produce the desired outcome. It therefore provides a medium for process planning, optimization, adaptation, and control.
Conditioning the Model on Process Physics
A manufacturing process cannot be modeled independently of the conditions under which the interaction occurs. The same robot motion can produce different outcomes depending on the material, tool, consumable state, contact conditions, and operating parameters. A useful Process World Model therefore needs to be explicitly conditioned on the variables that govern the process: St+1 = f(St, G, M, T, P, At), where G captures part geometry, M material properties, T the tool and consumable state, P the process conditions, and At the candidate process action. This physics-aware conditioning is what enables the model to generalize across manufacturing variation. A new part geometry, material, or tool configuration does not necessarily require learning the process from scratch. If the underlying interaction remains within a familiar physical domain, the model can adapt its prediction based on the new conditions rather than treating every configuration as an entirely new problem. The boundary of that generalization is set by the underlying process physics. Closely related processes such as sanding, grinding, and polishing may share substantial structure around contact, force, relative velocity, and material response. Processes governed by fundamentally different mechanisms may require different learned dynamics or additional state variables. The goal is therefore not a single model that treats every manufacturing process as identical, but a common modeling framework that generalizes where the physics is shared and adapts where it is not. Figure 1 summarizes this framework. The Process World Model is pretrained to learn the dynamics of the manufacturing process across a broad range of operating conditions using physics-based simulation and experimental data. At deployment, the model is not retrained from scratch for every new part or material. Instead, a small number of controlled interactions provide observations of the current process response. These measurements are converted into a physics-aware conditioning state that identifies the relevant process regime and anchors the model’s predictions to the physical system in front of it. Conditioned on this state, the world model predicts how candidate process actions will change the surface, enabling downstream planning to select process parameters i.e. recipes that achieve the desired manufacturing outcome (Refer to our article on CURIE, to understand how we leverage this model for recipe learning).
Learning Spatial Process Dynamics
Manufacturing processes such as surface finishing are inherently spatial. Different regions of a part experience different forces, orientations, velocities, and dwell times, producing a spatially varying process outcome. We represent the evolving workpiece state as a learned latent representation that captures geometry and surface condition. The world model then learns the conditional dynamics: p(St+1| St, At, Ct), where St is the current process state, At is the proposed tool interaction, and Ct contains the physics-aware conditioning variables describing the material, tool, contact, and process characteristics. A generative dynamics model predicts a distribution over future surface states rather than a single deterministic outcome. This allows the model to capture uncertainty arising from abrasive wear, material variation, contact compliance, and imperfect sensing. This probabilistic prediction is valuable for planning: candidate actions can be evaluated not only by their expected outcome, but also by the likelihood of satisfying process specifications without over-removal or surface damage.
Generalization Comes From Shared Physics
Generalization in manufacturing is constrained by the physics of the process. Sanding, grinding, and polishing differ in their operating characteristics and desired outcomes, but share important physical structure: contact mechanics, relative motion, applied force, material response, tool state, and geometry. A Process World Model can therefore reuse learned representations and dynamics across processes that share these underlying mechanisms. That generalization has boundaries. Processes governed by substantially different physics, such as spraying or welding, may require different state representations, conditioning variables, sensors, and learned dynamics. The scalable objective is therefore not a single universal model, but a reusable world-modeling framework: shared representations + physics-aware conditioning + learned dynamics + adaptation from physical observations. This enables reuse where physical structure is shared, while allowing the model to adapt as the underlying process physics changes.
Toward Robots That Understand Manufacturing
In factories of the future, robots need more than capable perception and motion. They need models that predict how physical actions transform the part. Process World Models provide this capability by representing manufacturing state, conditioning predictions on process physics, and learning how candidate interactions affect production outcomes. This allows robotic systems to reason about consequences before acting and to move beyond executing trajectories toward predicting, optimizing, and adapting manufacturing processes. That transition from understanding motion to understanding physical transformation, is a critical step toward Physical AI systems with true process intelligence.
Case Study: Sanding World Model
This case study focuses on our Sanding Process World Model (Refer Figure 2), an instantiation of the broader Process World Model architecture for surface finishing. The model takes the current surface state, its scanned geometry and measured condition together with a candidate sanding action, including path, force, abrasive grit, spindle speed, and dwell time. Conditioned on real observations of the specific tool-material-part interaction, it predicts the resulting material removal and surface finish.
The model is pretrained using data generated from a physics-based sanding simulator. The purpose of simulation is not to serve as ground truth, but to provide physics-grounded pretraining across a broad distribution of geometry, contact, tool, material, and process parameters. This exposes the model to the structure of sanding dynamics before it encounters a production part. Physics alone, however, is insufficient. Parameters such as pad compliance, abrasive behavior, wear, and material response are difficult to identify exactly, and prediction error in these quantities can propagate directly into the predicted outcome.
Pad stiffness provides a useful example. On a curved panel, stiffness determines how much of the sanding pad contacts the surface. Across physically plausible stiffness values, the simulator predicts contact over roughly 5% to 97% of the pad face. An overly stiff pad model concentrates removal around the high point of the surface and produces sharp boundaries where simulated contact terminates, while the physical coupon exhibits a smoother removal pattern. The simulator captures the governing structure of the interaction, but uncertainty in a single latent physical parameter can materially alter its prediction.
A neural model trained only on production data faces the inverse problem: it can capture complex process effects without explicitly identifying every physical parameter, but collecting enough real-world data to span the variation in materials, tools, geometries, and process settings is economically impractical. Physics provides scalable structure but is limited by uncertain parameters; neural models can capture effects that are difficult to model explicitly but cannot economically learn manufacturing physics from real data alone. Physics-aware conditioning connects the two. The Process World Model uses physics-grounded pretraining to establish a prior over process dynamics, then conditions those dynamics on real observations from the current tool–material–part interaction.
We evaluated this using a curved painted coupon and a controlled dwell experiment (Refer Figure 3). Twelve locations were sanded across a range of forces, spindle speeds, and dwell times, with coating thickness and roughness measured before and after each interaction. These input–outcome pairs form the conditioning context supplied to the model, allowing its latent dynamics to adapt to the observed response of the current material, tool, and geometry without retraining the model from scratch.
The clearest improvement is in surface roughness prediction (Refer Figure 4). Across 12 held-out sanding interactions comprising 36 Ra measurements, the observation-conditioned model achieved a root mean square error (RMSE) of 0.55 µm Ra, compared with 1.23 µm without conditioning and 1.21 µm with fitted physics-based conditioning, a roughly 55% reduction in prediction error. For material removal (Refer Figure 5), the neural variants produced similar average RMSE, but the figure below reveals an important distinction: the physics-based model predicts a biased contact pattern, producing a reasonable average removal while missing the true spatial contact dynamics. The neural model better captures where material is actually removed.
The important result is the division of labor between the components: simulation provides a physics-grounded prior, the neural dynamics model learns the high-dimensional process response, and real observations condition those dynamics to the current physical context. This allows the model to adapt rapidly without requiring the simulator to be perfectly identified or the neural model to relearn the process from physical data alone.
The architecture itself is not specific to sanding. The same Process World Model formulation can be instantiated for grinding, polishing, blasting, spraying, and other manufacturing processes. What changes are the state representation, action variables, conditioning context, and predicted outcome fields; the underlying architecture: physics-grounded pretraining + conditional learned dynamics + real-world context, remains the same.
