Industrial Robot Stiffness Optimization Boosts Milling Precision
A groundbreaking study from Chinese aerospace engineers has developed a new method to dramatically improve the precision of industrial robots used in high-stakes manufacturing, particularly for large spacecraft components. By optimizing the robot’s physical posture based on stiffness analysis, the research team achieved a 45% improvement in the flatness of machined surfaces, a critical metric for quality in aerospace applications. This innovation addresses a long-standing challenge in robotic machining: the inherent flexibility of robotic arms, which can lead to vibrations, tool deflection, and unacceptable levels of error during tasks like milling.
The research, spearheaded by a team from the China Academy of Space Technology and Beijing Satellite Manufacturing Plant Co., Ltd., focuses on a fundamental limitation of serial-type industrial robots. Unlike rigid, fixed-base CNC machines, robots are composed of a series of linked segments connected by rotating joints. While this grants them unparalleled flexibility and a large workspace, it also means the entire structure is more compliant, or less stiff. When a robot wields a milling tool and applies significant cutting forces to a workpiece, the force can cause the robot’s arm to bend and twist. This minute deformation at the robot’s base and joints translates into a much larger positional error at the tool tip, directly degrading the quality of the finished part. For the production of large, complex aerospace structures like satellite bodies or rocket fuselages, where tolerances are extremely tight, this error is a major barrier to adopting robotic systems for primary machining operations.
The team’s work, published in the journal Robot, tackles this problem at its root by shifting the focus from the robot’s inherent hardware limitations to its operational configuration. Instead of trying to build a stiffer robot, which would be costly and heavy, they asked a different question: For any given machining task, what is the single best way to position the robot’s arm—the specific angles of its six joints—so that it is as stiff and resistant to deformation as possible at that exact moment? This concept, known as “pose optimization,” leverages the fact that a robot’s stiffness is not a constant value; it varies dramatically depending on its geometry. A robot might be very stiff when its arm is extended straight out but much more flexible when it is folded into a complex, contorted pose.
The cornerstone of their approach is a sophisticated mathematical model that maps the robot’s internal joint stiffness to its overall performance at the tool tip in Cartesian space. The researchers began by establishing a comprehensive static stiffness model for the KUKA KR500 robot, a heavy-duty industrial model chosen for its ability to handle the forces involved in milling. They employed the principle of virtual work, a fundamental concept in mechanics, to derive the relationship between the forces applied at the end of the robot and the resulting deformations. This model showed that the robot’s overall stiffness at its end-effector is a complex function of two primary factors: the intrinsic stiffness of each of its six individual joints and the current configuration of those joints, which determines the robot’s Jacobian matrix—a mathematical representation of how joint movements translate into tool movements.
To make this model accurate and usable, the team had to determine the actual stiffness values of the robot’s joints, a parameter often not provided by manufacturers or difficult to calculate from first principles. They designed and conducted a series of identification experiments to measure these values empirically. In a controlled laboratory setting, they placed the robot in six different, carefully chosen poses. At each pose, they applied a series of known loads to the robot’s end flange using a system of pulleys and calibrated weights. Simultaneously, they used a high-precision six-axis force/torque sensor to measure the exact force being applied and a laser tracking system to measure the resulting tiny deformations of the robot’s arm with micron-level accuracy. By collecting this data across multiple poses and load levels, they created a large dataset that allowed them to solve an overdetermined system of equations. Using a least-squares method, they were able to back-calculate the individual stiffness of each joint, creating a complete and validated stiffness model for the KR500.
With this accurate model in hand, the next challenge was defining what “optimal stiffness” actually means for a milling task. The researchers recognized that a generic measure of overall stiffness is insufficient. Milling is a complex process where forces act in multiple directions, and the quality of the finished surface depends on how the robot resists these forces. To create a relevant and practical optimization goal, they developed a novel performance index based on the concept of the “compliance ellipsoid.” This is a geometric representation of the robot’s flexibility. Imagine a sphere of unit force applied from every direction to the robot’s tool tip. Due to the robot’s anisotropic stiffness—its varying resistance in different directions—this sphere is deformed into an ellipsoid. The shape and orientation of this ellipsoid reveal the directions in which the robot is most and least flexible.
For a milling operation, the researchers identified two critical aspects of this ellipsoid that dictate machining quality. The first is the stiffness within the plane of the cut. A high-quality milled surface requires consistent material removal. If the robot is much stiffer in one direction within the plane (e.g., forward) than in the perpendicular direction (e.g., sideways), it will deflect differently as the tool moves along its path, leading to variations in cut depth and a rough, uneven surface. To quantify this, they introduced a measure of “isotropy” for the ellipse formed by the intersection of the compliance ellipsoid and the milling plane. An isotropic ellipse is nearly circular, meaning the robot’s stiffness is nearly the same in all directions within the plane. A higher isotropy value indicates a more uniform stiffness, which allows for a smoother cut and potentially higher feed rates. They validated this by conducting milling tests on aluminum alloy with two different robot poses. The pose with the higher calculated isotropy produced a surface with significantly better roughness and flatness, confirming the metric’s practical relevance.
The second critical aspect is the stiffness in the direction normal to the milling plane. This is the direction of the primary cutting force as the tool plunges into the material. Any deflection in this direction directly changes the depth of the cut, which is a primary source of geometric error, such as poor flatness on a milled surface. They defined a “normal evaluation index” based on the length of the compliance ellipsoid’s semi-axis in this normal direction. A smaller value for this index indicates a stiffer robot in the direction that matters most for maintaining the correct cut depth. They validated this by loading the robot in different poses and measuring the actual deflection; the pose with the smallest normal index exhibited the least deflection under load.
The core of the research was the optimization process itself. The goal was to find the single robot pose—the set of six joint angles—that simultaneously maximized the in-plane isotropy and minimized the normal deflection index. This is a classic multi-objective optimization problem with multiple constraints. The constraints included the physical limits of each joint’s range of motion, the need to avoid self-collision of the robot’s links, and the absolute requirement that the robot’s tool tip must be positioned at the correct location and oriented with its axis perpendicular to the workpiece surface. Solving this complex problem required a powerful computational method.
The team chose a genetic algorithm, a type of evolutionary computation inspired by natural selection. Unlike traditional optimization methods that can get stuck in local minima and are highly sensitive to their starting point, genetic algorithms are robust global optimizers. They work by creating a population of potential solutions (robot poses), evaluating their “fitness” based on the optimization criteria, and then mimicking biological processes like selection, crossover, and mutation to evolve this population over many generations toward an optimal solution. The fitness function was carefully crafted to combine the two stiffness metrics, with a penalty applied to any solution that violated the position and orientation constraints. This “penalty function” approach ensures the algorithm searches for a solution that is both stiff and physically correct.
The results of the optimization were put to the ultimate test: real-world milling trials. The researchers used the optimized pose and compared it to a non-optimized, redundant pose—another configuration where the robot’s tool tip was in the exact same location and orientation. This is crucial because it isolates the effect of the arm’s configuration from other variables. They milled identical 80mm x 80mm flat surfaces on 5A06 aluminum alloy, a common aerospace material, using the same cutting parameters. The finished surfaces were then meticulously measured using a laser tracker. The data points from the milled surface were imported into analysis software, which used a least-squares algorithm to fit an ideal reference plane. The flatness was calculated as the distance between two parallel planes that just barely “sandwich” all the measured points—the smallest possible gap that contains the entire surface.
The difference was stark. The surface milled with the robot in the optimized pose had a flatness of 0.17 mm, while the surface milled in the non-optimized pose measured 0.31 mm. This represents a 45% improvement in a key quality metric, a result that far exceeds typical manufacturing tolerances and demonstrates a tangible, significant gain in performance. This experimental validation is the most compelling evidence for the method’s effectiveness.
This research has profound implications for the future of manufacturing, particularly in the aerospace industry. Large spacecraft components are often too big for conventional CNC machines, requiring expensive, custom-built equipment with long lead times. Mobile robotic systems, like the one used in this study which is mounted on an omnidirectional platform, offer a flexible and potentially more economical alternative. However, their adoption has been hampered by concerns over precision. This stiffness-based pose optimization method provides a software-driven solution to this hardware limitation. It allows manufacturers to get more performance out of existing robotic hardware, enabling them to machine large, complex parts with the quality required for flight.
The methodology is not limited to milling. The principles of stiffness modeling, empirical joint identification, and pose optimization based on task-specific performance indices could be adapted for other robotic processes like drilling, polishing, or even assembly, where force application and positional accuracy are critical. The use of a genetic algorithm also makes the approach highly adaptable; by changing the fitness function, the same framework could be used to optimize for other goals, such as minimizing energy consumption or maximizing dynamic performance.
In conclusion, the work by Chen Qintao, Yin Shen, Zhang Jiabo, and their colleagues represents a significant leap forward in robotic machining. By moving beyond the limitations of the robot’s physical design and intelligently optimizing its operational posture, they have demonstrated a practical and highly effective way to achieve a 45% improvement in machining precision. This research, published in Robot, provides a powerful new tool for manufacturers seeking to leverage the flexibility of industrial robots for high-accuracy tasks in aerospace and beyond. It is a testament to the power of combining fundamental mechanical principles with advanced computational techniques to solve real-world engineering challenges.
Chen Qintao, Yin Shen, Zhang Jiabo, Yue Yi, Zhou Yinghao, Wen Ke, Yang Jizhi, Bai Xiaopeng from China Academy of Space Technology and Beijing Satellite Manufacturing Plant Co., Ltd., Robot, DOI: 10.13973/j.cnki.robot.200042