Model Predictive Control (MPC) is the gold standard for controlling complex systems with strict safety constraints, such as drones, robotic arms, and power grids. Its power lies in its foresight: at every time step, it solves a mathematical optimization problem to find the best sequence of future actions. However, this foresight comes at a high price. Solving these non-linear optimization problems in real-time is computationally expensive, often requiring powerful processors and causing latency that makes high-frequency control impossible.
A growing body of research is solving this bottleneck by shifting the computational burden from the "online" control loop to an "offline" training phase. By training neural networks to mimic the behavior of expensive optimizers, researchers are achieving speedups of several orders of magnitude, enabling advanced control on low-power, resource-constrained hardware.
The Shift to Amortized Optimization
The core intuition behind this direction is amortized optimization. Instead of asking a robot to solve a difficult math problem 100 times per second, we solve thousands of variations of that problem in a factory setting (offline). We then use the resulting data—pairs of "system states" and "optimal actions"—to train a function approximator, typically a Deep Neural Network (DNN).
During deployment, the robot no longer "thinks" or optimizes; it simply performs a single forward pass through the network. This transforms a variable-time iterative search into a fixed-time matrix multiplication. Keyou You, Fei Dong, and Xingchen Li demonstrated this by implementing a DNN-based MPC on an FPGA for UAV tracking, achieving sub-millisecond latency—far faster than traditional onboard solvers Standoff Tracking Using DNN-Based MPC with Implementation on FPGA.
Distilling Complex Policies
This approach is often framed as "distillation" or "policy learning." A slow, "teacher" model (the MPC) generates optimal trajectories, and a fast, "student" model (the neural network) learns to replicate them.
- Recurrent Structures: Jingliang Duan and Shengbo Eben Li introduced Recurrent Model Predictive Control (RMPC), which uses recurrent neural networks to adaptively adjust prediction horizons, running five times faster than traditional solvers Recurrent Model Predictive Control.
- Data-Enabled Control: In the realm of model-free control, Xunyuan Yin and Xuewen Zhang developed Deep DeePC, which reduces computation time by 99% by training a network to predict the control operator directly from historical data Deep DeePC: Data-enabled predictive control with low or no online optimization using deep learning.
- Collective Dynamics: For swarms of hundreds of agents, Giacomo Albi and Dante Kalise use neural networks to bypass the "curse of dimensionality," enabling real-time feedback control that would otherwise be mathematically intractable Data/moment-driven approaches for fast predictive control of collective dynamics.
Bridging the Gap Between Speed and Safety
The primary critique of neural network controllers is that they are "black boxes" without the rigorous stability guarantees of classical MPC. Recent work has focused on "Safe-by-Design" approximations.
Johannes Köhler, Sebastian Trimpe, and Frank Allgöwer have pioneered methods to maintain statistical and deterministic safety guarantees. Their work on the KUKA robot arm achieved a 200x speedup while ensuring the robot never violated physical constraints Safe and Fast Tracking on a Robot Manipulator: Robust MPC and Neural Network Control. Similarly, the ALKIA-X algorithm uses kernel interpolation to provide a "one-shot" transformation of slow MPC into a fast explicit controller with closed-loop stability guarantees Automatic nonlinear MPC approximation with closed-loop guarantees.
Learning the Control Operator
A more radical perspective treats control not as a sequence of points, but as an operator learning task. Instead of learning a specific path, the model learns the mapping from "problem descriptions" (like goals and costs) to "solution trajectories."
Junchi Yan and Mingquan Feng developed the Neural Adaptive Spectral Method (NASM), which treats control as an operator. This method achieved a staggering 6,000x speedup over classical solvers, demonstrating that a well-trained network can generalize to entirely new scenarios it didn't see during training Optimal Control Operator Perspective and a Neural Adaptive Spectral Method.
Real-World Impact
This shift is enabling high-performance control in domains where it was previously impossible:
- Assistive Robotics: Koushil Sreenath and Ayush Agrawal used neural approximations to achieve hands-free dynamic walking for paraplegics using the ATALANTE exoskeleton Feedback Control of an Exoskeleton for Paraplegics: Toward Robustly Stable Hands-free Dynamic Walking.
- Maritime Autonomy: Rizhong Wang and Huiping Li applied policy learning to Unmanned Surface Vehicles (USVs), allowing complex nonlinear strategies to run at high frequencies on embedded hardware Policy Learning for Nonlinear Model Predictive Control with Application to USVs.
- Power Electronics: Xinyu Zhang 0007 and Zhigang Liu 0005 utilized pre-calculated lookup tables—a simpler form of offline approximation—to suppress high-frequency resonance in railway power systems The WRHE-PWM Strategy With Minimized THD to Suppress High-Frequency Resonance Instability in Railway Traction Power Supply System.
By moving the "intelligence" of the controller from real-time numerical solvers to pre-trained neural models, this research direction is making sophisticated, optimal control accessible to the smallest and fastest devices in the physical world.