In autonomous racing, swarm robotics, and agile manufacturing, the bottleneck is rarely the lack of a mathematical model. We have excellent physics-based frameworks like Model Predictive Control (MPC) that can navigate complex constraints and dynamics. The real enemy is latency.
Traditional optimal control requires solving a complex, non-convex optimization problem at every single time step. For a drone or a racing car, a "time step" might be only a few milliseconds. If the solver takes 50 milliseconds to converge, the vehicle has already traveled several meters with an outdated command, leading to instability or crashes.
To bridge this gap, researchers are developing techniques to "bypass" the online optimization bottleneck, moving from solving problems from scratch to high-speed predictive inference.
The Distillation of Expertise
The most direct way to eliminate optimization latency is to replace the solver with a neural network. In this paradigm, the optimization-based controller (like MPC) acts as an "offline expert." We generate thousands of optimal trajectories and then train a neural network to mimic the expert's output.
As Shouvik Roy and colleagues demonstrate in Neural Flocking: MPC-based Supervised Learning of Flocking Controllers, this distillation allows complex behaviors—like predator avoidance in swarms—to be executed in real-time using only local information. The intuition is simple: while solving a Quadratic Program (QP) is computationally expensive, a forward pass through a neural network is a series of highly optimized matrix multiplications. This can result in massive speedups, such as the 40x improvement reported by Michele Magno in the TC-Driver framework.
Unrolling Solvers into Neural Architectures
A common criticism of pure neural network controllers is that they are "black boxes" that lose the formal safety and constraint-handling guarantees of MPC. A middle ground is unrolling, where the structure of an optimization algorithm is embedded directly into the neural network's architecture.
In MPC-inspired reinforcement learning for verifiable model-free control, Yihan Zhou and team introduce "Learned Quadratic Programming" (LQP). Instead of a generic MLP, the network is structured to mimic the iterations of a solver. This provides a "best of both worlds" scenario: the speed of a neural network with the mathematical verifiability of a controller. Similarly, Hassan Iqbal explores Zero-Shot Function Encoder-Based Differentiable Predictive Control, which uses differentiable programming to allow controllers to adapt to new dynamics without the heavy lifting of traditional online re-optimization.
Real-Time Iteration and Sensitivity
If we must use an online solver, we don't necessarily need to wait for it to reach perfect convergence. The Real-Time Iteration (RTI) scheme is built on the intuition that a single, well-initialized iteration of a solver is often "good enough" to maintain stability, provided the control loop is fast enough.
Mario Zanon shows in Reinforcement Learning Based on Real-Time Iteration NMPC that by performing sensitivity analysis on these approximate solutions, we can still provide the gradients necessary for reinforcement learning. This allows agents to learn optimal policies in real-time without the prohibitive cost of full Nonlinear MPC (NMPC) convergence at every step.
Managing the Computational Budget
When hardware is extremely constrained, we can't just hope the solver finishes in time; we must actively govern it. Jordan Leung and Ilya Kolmanovsky proposed a computational governor that monitors the solver's progress. If the optimization is becoming too difficult (e.g., the vehicle is approaching a complex constraint), the governor modifies the reference command to steer the system back into a "comfort zone" where the solver can converge quickly.
This shift—from viewing optimization as a static math problem to viewing it as a dynamic, resource-constrained process—is what enables high-speed autonomy. Whether through the hierarchical approach of Dvij Kalaria in Local NMPC on Global Optimised Path for Autonomous Racing or the fast fault-estimation filters of Chris van der Ploeg, the goal remains the same: ensuring that the "optimal" action is also a "timely" one.