Behind the smooth and precise movements of an industrial robot lies a highly complex mechanical and computational system. Every movement of a robotic arm, from lifting and rotating to transferring objects, requires careful calculations to ensure that it follows the given instructions.

But do robots always move exactly as instructed?

In reality, not always. Friction between components, changes in load, mechanical characteristics, and various other disturbances can cause a robot’s movements to deviate slightly from the intended positions. Even when mathematical models are used to regulate its movements, a robot’s control system still faces limitations in predicting all the disturbances that may occur during operation.

This issue became the focus of researchers M. Sabri, M. Salamul Fajar Sabri, Faris Ahmad Mizanus Sabri, and Geubrina Hikmah in a study entitled State-space Feature Engineering for GPR-based Feedforward Compensation in a 5-DOF Robotic Manipulator, published in the scientific journal Array in 2026.

The study offers an interesting approach: using artificial intelligence to enable robots not only to move based on mathematical calculations but also to learn the characteristics of mechanical disturbances that have traditionally been difficult to predict.

Teaching Robots to Recognize Disturbances

The research was conducted using the Mitsubishi Movemaster RV-M1, an industrial robotic manipulator with five degrees of freedom, or five axes of movement.

Simply put, a robotic manipulator can be imagined as a human arm. Each section has joints that allow specific movements. To operate accurately, every joint must move in proper coordination. In conventional systems, robot movements are controlled using mathematical models that account for various factors, such as mass, gravity, position, velocity, and acceleration.

However, these models cannot always represent every real-world condition. One contributing factor is friction in the robot’s mechanical components. Friction can vary depending on movement speed, direction, and loading conditions.

As a result, a difference may arise between the torque required by the robot under actual operating conditions and the torque estimated by the mathematical model. This difference is known as residual torque.

To address this issue, the researchers used artificial intelligence to enable the system to learn disturbance patterns from robot movement data. They compared two approaches: Gaussian Process Regression (GPR) and Artificial Neural Network (ANN).

GPR is a machine learning method that not only generates predictions but also provides an indication of the uncertainty associated with those predictions. Meanwhile, ANN learns patterns and relationships among data through a computational network inspired by the way human nerve cells work. Both methods were tested to determine which approach was more effective in estimating mechanical disturbances in robots.

It Is Not Just About the Algorithm: How Data Is Interpreted Also Matters

One of the most interesting findings of this study is that the success of artificial intelligence depends not only on the algorithm being used but also on how data is presented to the system.

In the initial tests, the researchers used time as a reference for estimating disturbances experienced by the robot. This approach had a limitation. When the robot entered a movement period outside the time range used during training, the model’s predictive performance declined significantly.

The researchers then tried a different approach by using the robot’s motion states as the basis for learning. These states included the position, velocity, and acceleration of each robot joint.

The results were remarkable. Using motion-state information reduced prediction errors by up to 55.7% for the second joint and by more than 99% for the other four joints, compared with the time-based approach in the tests conducted.

These findings demonstrate that artificial intelligence systems perform better when provided with information directly related to the physical phenomena they are designed to learn.

In other words, it is not enough for a robot to know when a movement occurs. It also needs to recognize the motion state it is experiencing.

Furthermore, the study found that velocity was the dominant factor in explaining mechanical disturbances in four of the robot’s five joints. This is related to the characteristics of friction, which change with movement speed.

The second joint, however, exhibited more complex characteristics. Its disturbances were influenced not only by velocity but also by position, acceleration, and changes in motion state.

This finding also shows that different robot components may have distinct mechanical characteristics and therefore require appropriately tailored learning approaches.

GPR vs. ANN: Which Performs Better?

In the comparison between the two methods, GPR demonstrated better performance across most of the robot’s joints, particularly when mechanical disturbances followed patterns that were relatively consistent with velocity.

GPR’s advantage lies in its ability to learn relationships between input data and system responses while also providing information about prediction uncertainty.

However, the study also found that ANN could deliver competitive results. For the second joint, for example, an artificial neural network whose configuration had been optimized produced slightly better predictions than GPR.

This means that a sophisticated algorithm is not automatically the best choice for every problem. Selecting a method requires consideration of the data characteristics, system complexity, and type of disturbance to be predicted.

This principle is one of the study’s important contributions. Artificial intelligence is not simply about choosing the most advanced algorithm; it is about finding the right way to help an algorithm understand real-world problems.

Reducing the Workload of Robot Control Systems

Another important finding concerns the application of prediction results to compensate for robot movements. In conventional control systems, one widely used method is PID (Proportional-Integral-Derivative). Simply put, PID works by comparing a robot’s actual position with its desired position and then applying corrections when deviations occur.

In this study, GPR was used as a supporting system to estimate mechanical disturbances before additional corrections were required from the PID controller. In this way, GPR takes over part of the correction task that would otherwise have to be handled by PID.

In a closed-loop control system analysis using a robot model identified from experimental measurements, this approach showed promising results. Under load variations ranging from 0 to 1.5 kilograms, the use of GPR reduced the PID control effort by approximately 72–99% across the first four joints.

This reduction in workload indicates that the control system no longer needs to work as hard to correct disturbances that can already be predicted. The impact is reflected in the system’s improved ability to maintain movement accuracy while reducing the corrective effort required as the robot performs its tasks.

Nevertheless, these results should be interpreted in the proper context. The compensation tests were conducted through closed-loop control system analysis using an identified robot model, rather than through direct demonstrations of the compensator operating on a physical robot under all industrial conditions.

Toward More Adaptive Industrial Robots

This study illustrates that the future development of industrial robots cannot rely solely on improvements to mechanical components and conventional control systems.

Integrating artificial intelligence creates opportunities for robots to recognize their own operating characteristics, learn recurring disturbances, and respond more appropriately to changes in motion conditions.

Nevertheless, several challenges remain. The study shows that a model trained on movement in one direction may not necessarily predict disturbances accurately when the robot moves in the opposite direction. Similarly, changes in load and mechanical characteristics not covered by the training data remain issues that require further investigation. Future development will therefore need to incorporate a wider range of operating conditions, including variations in movement direction, load, and robot trajectories.

Ultimately, this study highlights an important point: the future of robotics will not be determined solely by how strongly or quickly a robot can move, but also by how well its internal systems understand its own movement behavior.

When mathematical models are combined with the learning capabilities of artificial intelligence, robots have the potential to operate with greater precision, adaptability, and efficiency. This provides a foundation for continued advances toward increasingly intelligent industrial automation systems.