As Machines Grow More Complex, Control Becomes Increasingly Critical
From simple actuation to fine-grained, computation-driven control
During my student years, my specialty was control engineering. After joining ITLAB, I conducted research on signal processing for millimeter-wave radar and LiDAR. Over the past decade, autonomous driving has gained momentum and control-related research challenges have increased, so I returned to the field of control. Now, beyond autonomous driving, I work on control for vehicle electrification centered on motion planning. Since control relies on observations, I believe my strength lies in knowing control and signal processing.
My background is in information theory, and I primarily conducted research on machine learning and other work that was done entirely on computers. After joining ITLAB, I worked on radar signal processing with sensing, and later on I began research on motor control. Until then, my research had focused on how to process acquired data, but I entered this field because I became interested in actuating physical systems based on dynamics.
In an era when mechanical structures were simple and requirements were not particularly demanding, control was not really considered important. It was sufficient for motors to rotate when powered on, and for vehicles to move when the accelerator was pressed. However, as machines gradually became more complex and the bar for various requirements rose such as improving ride comfort or suppressing unwanted vibration, control came to be emphasized as a discipline that demands fine-grained control using computers, far beyond simply “powering on and operating.”
The nature of control is also changing. Until about ten years ago, control meant moving objects according to mathematical formulas described in accordance with physical models. Recently, however, control has begun to see an AI-driven paradigm shift from hand-written equations based on physical models to data-driven approaches using machine learning and AI, similar to what we’ve seen in computer vision and natural language processing.
We maintain close, productive ties with the various business groups at headquarters
The Fascination and Fulfillment of Clear Success and Failure
Where physical motion makes success and failure explicit
What makes control fascinating is that objects actually move—making success and failure unmistakably clear. For example, in the case of autonomous driving brakes, control is responsible for stopping the vehicle precisely at the intended location. Since this is the final stage among many processes, and failure at this stage can invalidate the entire system, the responsibility is extremely significant—but the work is highly fulfilling.
I completely agree. I had been writing papers based on machine learning simulations for a long time, so I find it genuinely fascinating to see physical hardware actually move. Seeing it move is powerful evidence, so witnessing that right in front of me gives me a tremendous sense of fulfillment.
On the other hand, while control is often thought of as managing machine motion, I believe it is better understood in a broader sense as “planning to achieve objectives under uncertainty.” For example, designing manufacturing processes to be more efficient, improving component performance, or optimizing logistics—these themes can also be regarded as forms of control over longer time horizons.
This field is also seeing the introduction of machine learning and AI, and as computational performance improves, what we can accomplish continues to expand. People who maintain a broad perspective rather than being confined to the mobility domain may be able to do more interesting work.
Control Is Becoming More Relevant to Machine Learning and AI Researchers
Control converging with AI through data and computation
Because DENSO is a manufacturing company, control research has various applications not limited to automobiles themselves, including production technology and manufacturing processes. In a good way, we are close to the various business groups at headquarters, and there is a good foundation for collaborative progress. Even so, we are not pressed for short-term results in our research and can tackle themes over the long term. It’s a somewhat unusual environment, but I think researchers need the qualities to set long-term themes and manage them independently.
Personally, I want to see how control theory from the 1990s will evolve by 2030 through the integration of control and machine learning. Compared to LLMs and image-generation AI, which are evolving rapidly and seem poised to replace humans, fields such as robotics that actually move physical objects are in a sense the last frontier. A breakthrough in control comparable to the impact of LLMs is necessary, and I want to pioneer such a domain and take robotics to a point where it can truly replace human capabilities.
Compared to fields such as AI, machine learning, computer vision, and natural language processing, which are interconnected and where researchers can move between them easily, traditional control has been more separate as a field, and the two have traditionally been taught in separate university curricula”. However, in reality, the fields are rapidly converging. Once textbooks emerge, I think the integration of control and machine learning will accelerate all at once. When that happens, having the resources to gather data and compute at scale will matter more, so control research will likely trend in that direction.
Conversely, this also means that researchers who have worked on machine learning and AI until now can more easily enter the field of control. I’d love for them to join us and discover firsthand the satisfaction of making physical systems work as intended..
Join Our Research at DENSO ITLAB
Where engineering curiosity
meets real-world impact
