In control theory, Model Predictive Control (MPC) and Receding Horizon Control (RHC) are highly effective frameworks for managing complex systems with strict constraints. By solving an optimization problem over a finite future horizon at every time step, an MPC controller can safely guide a system—whether it is a drone navigating obstacles, a ship docking, or a robot arm manipulating objects.
However, when deployed in the real world, these controllers hit a major computational bottleneck: non-convexity. When system dynamics are non-linear or constraints are geometrically complex, the optimization landscape becomes non-convex. This means solvers can get trapped in local minima, and their execution times become highly unpredictable. For high-speed robotic systems or embedded hardware requiring millisecond-level response times, this computational latency can lead to instability or catastrophic failure.
To bridge the gap between theoretical control performance and real-time execution, researchers have developed several distinct paradigms to bypass or accelerate non-convex online optimization.
The Root of Non-Convexity in Real-Time Control
To understand why real-time constrained control is difficult, we must look at where non-convexity originates. In a standard control loop, the controller must find a sequence of inputs that minimizes a cost function while satisfying two main types of constraints:
- Dynamic Constraints: The physics governing how the system moves from one state to the next.
- Safety/Environmental Constraints: Boundaries the system must not cross, such as avoiding obstacles or keeping actuators within physical limits.
Non-convexity can emerge from either side. For example, as shown by Eugenio Cinquemani et al. in Convexity and convex approximations of discrete-time stochastic control problems with constraints, introducing probabilistic "chance constraints" to handle random disturbances typically warps a convex landscape into a non-convex one. Similarly, Debasish Chatterjee et al. demonstrated in Maximizing the probability of attaining a target prior to extinction that simply extending the time horizon in stochastic reach-avoid problems introduces non-convexity, making optimal policies mathematically intractable even for linear systems.
On the physical side, geometric constraints like collision avoidance are inherently non-convex; staying outside an obstacle means excluding a set from the search space, leaving a non-convex feasible region. As noted by Zhitao Wang et al. in Parallel Optimal Control for Cooperative Automation of Large-scale Connected Vehicles via ADMM, these geometric formulations create a severe trade-off between the precision of a safety model and the speed required for real-time execution.
Paradigm 1: Lifting Dynamics into Linear Latent Spaces
One of the most elegant ways to eliminate non-convexity arising from non-linear physics is to change the coordinate system. While a system's dynamics may look highly non-linear in its natural state space, Koopman operator theory proves that any non-linear dynamical system can be represented as a linear system if it is mapped (or "lifted") into an infinite-dimensional space of observable functions.
By approximating a finite-dimensional version of this lifted space, researchers can transform a non-convex nonlinear MPC problem into a standard convex Quadratic Program (QP).
Non-linear State Space (Non-Convex Optimization)
│
▼ [Lift via Koopman Operator]
Higher-Dimensional Latent Space (Linear Dynamics -> Convex Optimization)
This approach has been successfully applied across multiple domains:
- Medical Robotics: Zhichao Liang et al. utilized a Koopman-MPC framework in Online Learning Koopman operator for closed-loop electrical neurostimulation in epilepsy to linearize chaotic brain activity, enabling millisecond-level calculations for electrical stimulation pulses.
- Robust Control: Xinglong Zhang et al. combined data-driven Koopman modeling with tube-based robust control in Robust tube-based model predictive control with Koopman operators, achieving the computational efficiency of linear MPC while guaranteeing stability against modeling errors.
- Sequence Modeling: Zhaoyang Li introduced MamKO: Mamba-based Koopman operator for modeling and predictive control, which dynamically generates Koopman operators using the Mamba architecture to maintain linear control tractability in time-varying systems.
By ensuring the prediction model is linear, solvers can find global optima reliably and rapidly, satisfying the strict timing constraints of real-time hardware.
Paradigm 2: Convex Reformulations and Operator Splitting
When non-convexity stems from constraints rather than dynamics, researchers often use mathematical reformulations to break the problem into smaller, convex pieces or project them into structured mathematical cones.
Operator Splitting (ADMM and Alternating Minimization)
If an optimization problem contains a mix of smooth trajectory costs and non-smooth, non-convex obstacle constraints, it can be split. The Alternating Direction Method of Multipliers (ADMM) and Alternating Minimization (AM) algorithms allow agents to solve the trajectory optimization and the constraint satisfaction steps independently and iteratively.
For instance, in decentralized swarm robotics, Vivek K. Adajania et al. developed AMSwarm: An Alternating Minimization Approach for Safe Motion Planning of Quadrotor Swarms in Cluttered Environments. By using a polar reformulation of collision constraints combined with an AM solver, they decoupled the multi-agent optimization problem, enabling quadrotor swarms to navigate cluttered environments up to 42 times faster than standard baselines.
Conic Optimization
Another approach is to relax or reformulate complex constraints into Second-Order Cone Programs (SOCPs). SOCPs are a class of convex optimization problems that can be solved in polynomial time with highly optimized interior-point methods.
In Safe and Stable Control Synthesis for Uncertain System Models via Distributionally Robust Optimization, Kehan Long et al. addressed the non-convexity of probabilistic chance constraints by reformulating a distributionally robust safety framework into an efficient SOCP, allowing high-speed robots to maintain safety guarantees with minimal latency.
Paradigm 3: Shifting the Computational Burden Offline
If an optimization problem is too complex to simplify or linearize, another strategy is to change when the computation happens. Instead of forcing an onboard processor to solve a non-convex optimization problem online every few milliseconds, the problem can be solved thousands of times offline across a wide range of states. A function approximator, such as a Deep Neural Network (DNN), is then trained to learn the mapping from the system state to the optimal control action.
OFFLINE TRAINING:
Explore State Space ──> Solve Expensive Non-Convex MPC ──> Train Neural Network Policy
ONLINE RUNTIME:
Current State ──> Evaluate Neural Network (Matrix Multiplications) ──> Microsecond Control Action
This strategy underpins the work by Rizhong Wang et al. in Policy Learning for Nonlinear Model Predictive Control with Application to USVs. Their framework, PL-MPC, shifts the heavy computational lifting of Nonlinear MPC to an offline training phase, enabling high-precision control of Unmanned Surface Vehicles at runtime. Similarly, D. K. Siong Tok et al. proposed a factorized radial basis function model in Factorized f-step radial basis function model for model predictive control to eliminate the error accumulation of recursive neural network predictors while minimizing online numerical calculations.
Alternatively, instead of a neural network, the offline computational burden can be structured into a discrete bank of pre-computed linear controllers. As demonstrated by Babak Mehdizadeh Gavgani et al. in Real-Time Trajectory Adaptation for Tossing Robots Using Soft Switching Multiple Model Predictive Control, switching dynamically between a bank of optimized linear controllers yields a 1300x reduction in computational load compared to traditional nonlinear methods, allowing robotic arms to adapt trajectories mid-motion.
Paradigm 4: Decoupling Performance from Safety via Filtering
A final paradigm separates the control problem into two distinct layers: a high-level performance layer and a low-level safety layer. The high-level layer focuses on complex goals (like path planning or reinforcement learning) and can afford to be slow or unconstrained. The low-level layer, known as a Safety Filter or Reference Governor, acts as a supervisor. It takes the proposed command and minimally modifies it only if it threatens to violate a safety boundary.
By using Control Barrier Functions (CBFs), the safety filter can evaluate set invariance using a localized, linear inequality. This reduces the safety enforcement step to a simple, convex Quadratic Program that can be solved in microseconds.
┌──────────────────────────────────┐
│ High-Level Controller (RL / NMPC)│ ──> Desired Action (Unsafe/Aggressive)
└──────────────────────────────────┘
│
▼
┌──────────────────────────────────┐
│ Convex Safety Filter (CBF / QP) │ ──> Projected Safe Action to Actuators
└──────────────────────────────────┘
Key developments in this area include:
- Output-Feedback Filters: Lukas Brunke et al. introduced the Robust Predictive Output-Feedback Safety Filter for Uncertain Nonlinear Control Systems, which wraps around arbitrary controllers to guarantee safety even when state observations are noisy and incomplete.
- Differentiable CBFs: Wei Xiao developed a framework in Safe Neural Control for Non-Affine Control Systems with Differentiable Control Barrier Functions that integrates differentiable CBFs directly into neural ordinary differential equations, achieving safety verification speeds over 30 times faster than traditional optimization loops.
- High-Order Constraints: Shuo Liu proposed an iterative convex optimization method in Iterative Convex Optimization for Model Predictive Control with Discrete-Time High-Order Control Barrier Functions to handle complex, high-order safety constraints without incurring the high computational costs that typically prevent real-time robotic deployment.
By isolating safety constraints from global trajectory generation, these filtering techniques allow systems to leverage advanced, non-convex optimization or machine learning models at a high level while maintaining strict, real-time safety guarantees at the hardware level.
Key Foundational Work
Generalized predictive control of linear systems with actuator arrearage faults (2008) — contributed 1 concept elements to this direction.
Stochastic Receding Horizon Control With Bounded Control Inputs: A Vector Space Approach (2011) — contributed 1 concept elements to this direction.
Regret Bounds for the Adaptive Control of Linear Quadratic Systems (2011) — contributed 1 concept elements to this direction.