Real-Time Preview Control of Active Suspensions Under Actuation Limits via Learned MPC Surrogates and Online Identification

Every vehicle that travels over a road encounters a constant, jarring reality: the ground is rarely flat. When a car hits a bump, the wheels move up and down, and without intervention, that motion transfers directly to the passengers, creating a rough, uncomfortable ride. For decades, engineers have tried to solve this with passive systems, using rubber pads and mechanical springs that absorb some of the shock. Later, semi-active systems emerged, which could adjust their stiffness or damping to react to the road, but these still relied on the car's reaction to a bump after it had already been hit. The ultimate goal has always been an active suspension system, a setup that uses motors and sensors to push the car body in the opposite direction of a bump, effectively canceling out the vibration before the passengers even feel it. While this technology promises a smoother ride and better safety, it has remained a luxury feature found only on the most expensive cars. The primary barrier has been a difficult trade-off: the computers required to calculate the perfect counter-movement in real-time are too slow for standard car chips, and the motors needed to generate enough force are often too large or energy-hungry for mass production.

A team of researchers at Ben-Gurion University of the Negev and the Guangdong Technion-Israel Institute of Technology has proposed a new way to bridge this gap. They developed a control system that combines the mathematical precision of advanced planning with the speed of a learned neural network, all while respecting the physical limits of the car's motors. Their work, tested through detailed computer simulations, suggests that it is possible to achieve the smoothness of a perfect, ideal suspension without needing supercomputers or infinite power. The key insight is that the system does not just react to the road; it looks ahead. By using a "preview" of the road surface—data about bumps that are still a few meters away—the controller can prepare its movements in advance. However, the researchers found a critical nuance: this preview capability is most beneficial precisely when the motors are weak or saturated. When actuators are strong, they can simply react to disturbances, but when they are limited, the ability to look ahead allows the system to reallocate its limited force earlier in time, making the preview information far more valuable than in unconstrained scenarios.

To make this system practical, the researchers tackled the problem of speed. The most accurate way to control the suspension is through a method called model predictive control, which constantly calculates the best possible action for the next few seconds. In a perfect world, this calculation would be instantaneous, but in reality, it takes too long for a car traveling at highway speeds. The researchers solved this by training a neural network, a type of artificial intelligence, to mimic the decisions of the slow, perfect calculator. They fed the neural network thousands of examples of the ideal controller's actions, teaching it to recognize the road conditions and the car's state and respond with the same precision. Once trained, this neural network acts as a "surrogate," making decisions nearly eighty times faster than the original method. This speedup is significant enough that the system could run on the standard computer chips found in modern vehicles, rather than requiring specialized, expensive hardware.

The study also addressed a common problem in real-world driving: the weight of a car changes. A vehicle with a full load of passengers and luggage is heavier than one that is empty, and this change alters how the suspension behaves. A controller designed for an empty car will perform poorly when the car is full. To handle this, the team added an online identification module that continuously estimates the current weight of the vehicle while it is driving. By feeding this updated information into the neural network, the system adapts its strategy on the fly. In their simulations, this adaptive approach reduced the vibration felt by the passengers by an additional thirty-two percent compared to a system that assumed the weight never changed.

The researchers tested their method against a standard baseline using various road conditions, from smooth surfaces to rough, uneven terrain and even specific obstacles like speed bumps. They simulated a quarter of a vehicle, focusing on the vertical motion of a single wheel and the body above it, and later expanded the test to a half-vehicle model that included the pitching motion of the entire car. The results showed that the new system consistently outperformed traditional methods. When the motors had enough power, the preview-enabled system reduced the vibration acceleration by twelve percent compared to a standard feedforward controller. In the specific case of hitting a speed bump, the improvement was even more dramatic; with a 0.7-second preview window, the system achieved a 57% reduction in vibration acceleration compared to the baseline. The system worked by starting to push the car body upward before the wheel even touched the bump, a proactive move that prevented the motor from getting overwhelmed.

However, the study also revealed a fundamental boundary. The researchers found that the benefits of looking ahead at the road are maximized when the motors are weak and operating near their limits. When the force required to smooth the ride exceeds the motor's maximum limit, the system saturates; in these constrained conditions, the extra information about the road ahead becomes crucial because it allows the controller to optimize the use of its limited force. Conversely, if the actuators are extremely strong and never saturate, the advantage of the preview diminishes. This finding provides a clear guideline for engineers: to make preview control effective, the hardware does not need to be infinitely powerful, but the control strategy must be designed to leverage preview specifically to overcome actuator limitations. The team also discovered that looking too far ahead does not help; a preview window of about half a second was sufficient to capture the necessary information for the frequencies of road bumps they tested. Extending the view beyond this point increased the computational load without improving the ride.

The paper concludes that while the method has not yet been tested on a physical vehicle, the simulation results are robust and point toward a viable path for mass adoption. The approach offers a way to get the performance of a complex, optimal control system without the computational cost, making active suspension a more realistic option for everyday cars. By combining the foresight of preview control, the speed of a learned neural network, and the adaptability of real-time weight detection, the researchers have demonstrated a system that could one day make every ride feel as smooth as a luxury sedan, regardless of the road conditions or the number of passengers.