Robust Optimal Tracking Control for Center-of-Mass-Offset Mecanum Mobile Robots Using Off-Policy Integral Reinforcement Learning

In the world of robotics, movement is often a matter of trade-offs. A standard wheeled robot moves forward and turns, but to move sideways, it must first turn its entire body. This limitation makes navigating tight spaces, like crowded warehouses or narrow factory aisles, a slow and clumsy affair. To solve this, engineers developed the Mecanum wheel. These are standard wheels fitted with a ring of small, angled rollers. This simple addition allows the robot to slide in any direction—forward, backward, sideways, or even diagonally—without ever changing which way its front is facing. It is a powerful tool for automation, enabling machines to weave through obstacles with the agility of a person walking through a busy crowd.

However, making these robots move precisely is far more difficult than it appears. The physics of a Mecanum robot are complicated by two main factors. First, the wheels themselves are prone to slipping; the rollers do not grip the ground perfectly, leading to small, unpredictable errors in movement. Second, and perhaps more critical, is the issue of balance. When a robot carries a load, such as a box or a robotic arm, the center of its weight often shifts away from its geometric middle. This shift changes how the robot resists turning and how the weight is distributed across its four wheels. If a control system does not account for this shifting weight and the inevitable wheel slippage, the robot will drift off course, oscillate, or fail to follow its intended path.

Researchers at the Electric Power University in Hanoi, Vietnam, have tackled this problem by developing a new way to teach these robots how to stay on track. Their work focuses on a specific type of intelligent control that learns from experience rather than relying on a perfect, pre-calculated map of the robot's physics. In their study, they created a system that allows a four-wheeled Mecanum robot to track a desired path with high precision, even when the robot's weight distribution is changing constantly and the wheels are slipping. They achieved this by using a method called off-policy integral reinforcement learning, a technique that allows the robot to learn the best way to move by observing data collected during operation, rather than needing to know every detail of its own mechanics beforehand.

The core challenge the researchers addressed is that real-world robots are messy. Theoretical models assume perfect conditions, but in reality, a robot might carry a heavy load that shifts as it moves, or its wheels might lose traction on a smooth floor. Traditional control methods often struggle with these uncertainties, requiring engineers to tune the system conservatively or to have exact knowledge of the robot's weight and friction at all times. The new approach proposed in this paper treats the control problem as a game between two opposing forces. On one side is the robot's controller, which wants to minimize the effort needed to stay on the path. On the other side is the "worst-case" disturbance, representing the maximum possible error caused by slipping or shifting weight. The goal is to find a strategy that works best even when the environment is trying to push the robot off course as hard as it possibly can.

To solve this, the researchers used a mathematical framework known as the Hamilton–Jacobi–Isaacs equation. While this equation is too complex to solve with a simple formula for a real robot, the team used a clever workaround involving artificial intelligence. They employed three neural networks, which are computer systems modeled after the human brain, to learn the solution. One network acts as a critic, evaluating how well the robot is doing. A second network, the control actor, decides what motor commands to send to the wheels. The third network, the disturbance actor, learns to predict the worst possible errors the environment could throw at the robot. By training these networks together using data collected from the robot's movements, the system learns a robust strategy that balances smooth movement with the ability to resist disturbances.

The researchers tested their new controller in two distinct scenarios using computer simulations. In the first scenario, they simulated a robot with a fixed, unchanging weight offset, similar to a robot carrying a heavy box that never moves. In this stable environment, the new controller performed just as well as a standard learning method, keeping the robot on its circular path with high accuracy. The robot's motors produced smooth, steady torque, and the system learned quickly without becoming unstable. This confirmed that the new method works even when the robot's physical properties are constant but different from the ideal design.

The second scenario was far more demanding. Here, the researchers simulated a robot where the center of mass was constantly shifting, as if the robot were carrying a robotic arm that was moving its own weight around. They also introduced time-varying wheel slippage and changes in the total weight of the robot, mimicking a machine picking up and dropping off heavy loads. In this chaotic environment, the standard learning method began to struggle, showing larger errors in the robot's heading and less smooth control. In contrast, the new off-policy controller maintained a tight grip on the desired path. It successfully compensated for the shifting weight and the slipping wheels, keeping the tracking errors small and the motor commands smooth. The simulation showed that the new controller reduced the error in the robot's heading angle more effectively than the standard method, proving its superior ability to handle complex, changing uncertainties.

The study concludes that this new approach offers a significant improvement for mobile robots operating in unpredictable environments. By explicitly accounting for the worst-case disturbances and learning from data rather than relying on perfect models, the system achieves a level of robustness that traditional methods lack. The researchers verified that the system remains stable and that the learning process does not lead to erratic behavior, even under severe uncertainty. While the results so far are based on computer simulations, they provide a strong foundation for future work. The next step will be to test this controller on a physical robot, moving from the safety of a digital simulation to the messy, unpredictable reality of a factory floor. If successful, this technology could enable Mecanum robots to operate more autonomously and reliably in dynamic settings, from automated warehouses to service environments where conditions change constantly.