Smart and Simple: Model-Based Optimal Control for CNH

Designing intelligent control systems that can anticipate and adapt to changing conditions is a key challenge in automation. For complex cases, Model Predictive Control (MPC) is often seen as the ideal approach because it optimizes actions based on a model of the system and its predicted future behaviour. However, MPC is computationally demanding and not always practical for use in industrial machines.

That is why we developed a method to create simplified controllers that achieve performance close to MPC, but are much easier to design, implement, and tune. This enables industrial partners to apply powerful novel controllers efficiently and reliably in their own systems.

From Complex Models to Practical Control

The new approach starts with an MPC that provides the optimal control behaviour for a system. Using an automated design process, we then generate a lightweight version built from modular control blocks. These reproduce the key features of the original MPC but require far less computing power.

The result is a controller that retains the advantages of MPC- anticipating future events, respecting constraints, and coordinating subsystems - while remaining simple enough for practical use and quick retuning. This reduces the need for lengthy manual design cycles and allows faster adaptation to new operating conditions.

Assistant Schema

CNH Case: Following the Swath Line

To test this approach, we collaborated with CNH Industrial and applied it to a tractor-baler system tasked with automatically following the crop swath. This is a demanding control problem: while automating a single vehicle can be done relying only on classical PID control, here we need more advanced control to steer the tractor such that the baler trailing the tractor remains precisely aligned with the swath, despite bends, uneven ground or changing speed.

Previously, this required a manually designed controller that took significant effort to design and calibrate. With the new method, an equivalent controller could be generated automatically, achieving similar or better tracking performance, nearly on par with the MPC benchmark. And to achieve this, it only requires roughly the same computational load as a PID controller. The new controller also handles disturbances such as slip or steering delay effectively, reducing operator workload and improving consistency in baling.

Performance Visualised

Shown below are a set of example maneuvres as executed with the resulting controller. It can be seen the new controller can perform these nearly perfectly.

If we then check how the control signals of the new controller match the MPC signals, and how the resulting motion resembles that with an MPC, we see that both are very similar indeed, and thus that the new controller very closely approximates the MPC.

Beyond Agriculture

Although this case considered agriculture, this approach can benefit many sectors where systems are complex or predictive coordination is needed - for example, in manufacturing, mobility, or energy management. By simplifying model-based control design, we help bridge the gap between advanced algorithms and industrial deployment, supporting the move towards smarter, more autonomous machines that are easier to implement and maintain.