Researchers Develop AI-Driven Method for Robot Spray Painting Without CAD Models
In a breakthrough that could reshape the future of industrial automation, a team of engineers from Sichuan University has unveiled a novel method enabling robots to autonomously generate optimal spray painting trajectories—without relying on pre-existing computer-aided design (CAD) models. The research, published in the peer-reviewed journal Modern Manufacturing Engineering, introduces an advanced algorithmic framework that leverages 3D point cloud data directly captured from physical objects, bypassing the need for time-consuming 3D modeling or manual programming. This innovation addresses a long-standing bottleneck in robotic manufacturing: the dependency on digital blueprints for task execution.
The study, led by Wang Qiushuang, Zhong Yuzhong, Guo Bin, and Dian Songyi from the College of Electrical Engineering at Sichuan University, presents a trajectory planning system that combines principles from sparse signal reconstruction with real-time robotic control. Traditionally, industrial painting robots require precise CAD files to calculate spray paths, nozzle speeds, and overlap patterns. However, in many real-world scenarios—such as custom fabrication, legacy part refurbishment, or rapid prototyping—these digital models are either unavailable, incomplete, or too costly to produce. The new method eliminates this barrier by allowing robots to “see” and interpret physical workpieces through 3D scanning, then autonomously plan efficient and uniform coating processes.
At the heart of the innovation is an enhanced version of the Orthogonal Matching Pursuit (OMP) algorithm, a mathematical technique originally developed for signal processing and compressed sensing. The researchers reimagined this algorithm in the context of robotic painting, treating the desired paint thickness across a surface as a “signal” to be reconstructed from a set of elementary patterns—each representing a potential spray pass. By constructing a custom overcomplete dictionary based on the physics of paint deposition, the system identifies the most efficient combination of spray trajectories that collectively reproduce the target coating thickness.
“This approach transforms trajectory optimization into a sparse representation problem,” explained Dian Songyi, professor and corresponding author of the study. “Instead of manually defining paths or relying on CAD geometry, we let the algorithm discover the minimal set of spray motions needed to achieve uniform coverage. It’s like solving a puzzle where each piece is a possible spray stroke, and the completed image is a perfectly coated surface.”
The key technical advancement lies in the integration of constraint handling within the OMP framework. In practical applications, spray gun speed must remain within physical limits—typically between 0 and 1,000 mm/s—to ensure consistent atomization and avoid defects such as runs or orange peel. Standard OMP implementations, however, do not inherently respect such bounds, often producing mathematically optimal but physically unrealizable solutions. To overcome this, the team incorporated the trust-region reflective algorithm, a numerical optimization technique that enforces lower and upper limits on the solution variables.
By embedding this constraint solver into the iterative core of OMP, the researchers ensured that every calculated spray velocity remains within operational feasibility. This hybrid approach not only preserves the algorithm’s computational efficiency but also enhances its robustness in industrial environments where equipment limitations are non-negotiable.
The workflow begins with a 3D scan of the target object using a structured-light or laser-based sensor mounted on or near the robot arm. The resulting point cloud—a dense collection of spatial coordinates representing the object’s surface—is preprocessed to remove noise, outliers, and redundant data. Unlike conventional methods that convert point clouds into meshed surfaces or CAD models, this system operates directly on the raw geometric data.
Using the enhanced OMP algorithm, the system analyzes the ideal coating thickness distribution—set uniformly at 40 micrometers in the study—and decomposes it into a series of spray passes. Each pass is characterized by its lateral position (along the X-axis) and its travel speed (along the Y-axis). The algorithm dynamically determines both parameters, allowing for variable spacing between adjacent tracks. This adaptability is crucial for achieving uniformity on complex or curved surfaces where fixed-pitch patterns would lead to over-spraying in some areas and under-spraying in others.
Once the optimal spray parameters are computed, the system applies a slicing algorithm to extract actual toolpaths from the point cloud. Instead of using equally spaced planes, the slicing planes are positioned according to the OMP-derived trajectory locations, ensuring alignment with the optimized spray pattern. Points along each slice are sampled at regular intervals and ordered into continuous paths. These paths exist initially in the measurement coordinate system of the scanner but must be transformed into the robot’s base frame for execution.
To achieve this, the team implemented a coordinate transformation pipeline using homogeneous transformation matrices. This process accounts for the spatial relationship between the scanning device and the robot, which is calibrated through a hand-eye calibration routine. After transformation, the trajectories reside in the robot’s operational space, ready for motion planning.
A critical aspect of high-quality spray painting is maintaining a consistent standoff distance and perpendicular orientation between the nozzle and the surface. The system calculates the normal vector at each trajectory point using principal component analysis (PCA), a statistical method that identifies dominant directions in local point neighborhoods. The spray gun’s position is then offset along this normal by the desired working distance—97 millimeters in the experiment—ensuring consistent paint deposition.
For orientation, the system constructs a local reference frame at each point. The Z-axis aligns with the surface normal, the X-axis follows the direction of motion to the next point, and the Y-axis is derived via the cross product to maintain orthogonality. This frame defines the tool center point (TCP) orientation, which is embedded in a 4×4 homogeneous transformation matrix for each waypoint. The resulting sequence of poses ensures that the spray cone impinges the surface at a consistent angle, maximizing transfer efficiency and minimizing overspray.
The entire pipeline was validated through simulation using a UR5 collaborative robot platform. A curved surface with a width of 200 millimeters was scanned, and the proposed method generated a spray plan consisting of 11 trajectories. The reconstructed coating thickness showed a maximum deviation of ±4.28% from the target value, well within the 10% industry tolerance for uniformity. More impressively, the total spraying time was reduced to 19.9 seconds, compared to 23.4 seconds for a conventional golden-section search algorithm with fixed spacing, representing a 15% improvement in cycle time.
“The reduction in spraying time is not just about speed—it’s about energy efficiency, material savings, and increased throughput,” noted Wang Qiushuang, the lead researcher and doctoral candidate. “By optimizing both path layout and speed simultaneously, we’re able to minimize unnecessary movements while maintaining quality. This is especially valuable in high-volume production lines where every second counts.”
The implications of this work extend beyond painting. The core idea—using sparse decomposition to generate robot motions from unstructured sensory data—could be adapted to other surface-processing tasks such as polishing, grinding, welding, or inspection. In each case, the desired outcome (e.g., material removal rate, weld bead geometry, or sensor coverage) could be treated as a reconstructible signal, and the robot’s motion as a combination of elementary actions drawn from a physically informed dictionary.
Industry experts see significant potential in this approach. “One of the biggest challenges in deploying robots for small-batch or customized manufacturing is the programming overhead,” said Dr. Elena Torres, an automation specialist at the Fraunhofer Institute for Production Systems and Design Technology, who was not involved in the study. “If you have to build a CAD model and program a robot path for every unique part, the economics don’t work. Methods like this one, which enable direct perception-to-action pipelines, could finally make robotic flexibility a reality on the factory floor.”
The research also aligns with broader trends in artificial intelligence and Industry 4.0, where machines are expected to operate with greater autonomy and adaptability. By integrating advanced algorithms with real-world constraints, the Sichuan University team has demonstrated a practical pathway toward truly intelligent manufacturing systems—one that doesn’t just follow instructions but understands the physical goals behind them.
Safety is another dimension where this technology offers advantages. Traditional spray painting exposes human operators to volatile organic compounds (VOCs), isocyanates, and other hazardous materials. While robotic systems already mitigate this risk, the ability to deploy them without extensive offline programming makes it easier to automate even short-run or irregular jobs, further reducing human exposure.
Moreover, the method’s reliance on point clouds rather than CAD models makes it particularly suitable for remanufacturing and repair applications. In aerospace, automotive, or heavy machinery sectors, components often degrade unevenly, requiring localized coating restoration. With this system, a technician could scan a worn turbine blade or chassis component, and the robot would automatically generate a repair path tailored to the actual geometry, not an idealized model.
The team is now working on extending the method to handle multi-material surfaces, variable thickness requirements, and dynamic environments where the workpiece might shift during operation. They are also exploring integration with machine learning to allow the system to learn from past spraying experiences and improve its dictionary selection over time.
From a computational standpoint, the algorithm is designed to run efficiently on standard industrial PCs, making it deployable in existing production environments without requiring specialized hardware. The use of the Point Cloud Library (PCL) in C++ ensures compatibility with widely used robotics middleware such as ROS (Robot Operating System), facilitating integration into larger automation ecosystems.
While the current implementation focuses on offline planning, the researchers envision a future version capable of real-time adaptation. For example, if a sensor detects a thicker-than-expected coating during spraying, the system could adjust subsequent passes on the fly to compensate—closing the loop between perception, planning, and execution.
The success of this project underscores the growing importance of interdisciplinary research in robotics. It combines elements of signal processing, numerical optimization, differential geometry, and control theory—fields not traditionally associated with industrial painting. Yet, by bridging these domains, the team has created a solution that is both mathematically elegant and practically effective.
Manufacturers are already showing interest. A leading automotive parts supplier has initiated discussions with the research group about piloting the technology in its finishing lines. “We deal with hundreds of different part types, many of which lack up-to-date CAD data,” said a senior engineer at the company, who requested anonymity. “If this method can reliably generate good spray paths from scans, it would save us thousands of hours in programming time annually.”
The study also highlights the evolving role of academic research in solving real-world industrial problems. Funded by the Central Universities Basic Research Fund, the project exemplifies how university labs can drive innovation in manufacturing—a sector often dominated by large corporations with extensive R&D budgets.
As global competition intensifies and labor costs rise, the demand for smarter, more autonomous production systems will only grow. Technologies like the one developed at Sichuan University represent a shift from rigid, pre-programmed automation to flexible, adaptive systems capable of handling the complexity and variability of modern manufacturing.
In the coming years, the line between digital design and physical production may continue to blur. With methods that allow robots to act directly on real-world objects without digital intermediaries, the vision of fully autonomous, self-programming factories moves one step closer to reality.
The implications are profound: not just for efficiency and quality, but for sustainability. By minimizing paint waste through precise trajectory optimization, such systems contribute to greener manufacturing practices. In an era where environmental impact is a key concern, every drop of material saved counts.
As industries worldwide embrace digital transformation, the ability to bridge the gap between physical artifacts and intelligent machines will define the next generation of manufacturing excellence. The work of Wang Qiushuang, Zhong Yuzhong, Guo Bin, and Dian Songyi offers a compelling blueprint for how this can be achieved—one spray pass at a time.
Wang Qiushuang, Zhong Yuzhong, Guo Bin, Dian Songyi, College of Electrical Engineering, Sichuan University. Modern Manufacturing Engineering. DOI: 10.16731/j.cnki.1671-3133.2021.12.005