Computationally Tractable Robust Nonlinear Model Predictive Control using DC Programming

arXiv:2602.01164 · math.OC, cs.SY, eess.SY · Submitted 2026-02-01 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.

Rosa: Today's paper: "Computationally Tractable Robust Nonlinear Model Predictive Control using DC Programming".

Dev: This paper proposes a computationally tractable robust Model Predictive Control (MPC) framework for nonlinear systems by leveraging difference-of-convex (DC) programming and sequential convex programming.

Rosa: First, who's behind it and why it matters.

Title and authors: Rosa: So we're looking at the paper "Computationally Tractable Robust Nonlinear Model Predictive Control using DC Programming." It seems like they've tackled the difficulty of making robust nonlinear MPC computationally feasible by using difference-of-convex programming and sequential convex programming.

Dev: That sounds like they are aiming to solve a problem that usually has too much computational overhead or leads to overly cautious control designs in nonlinear MPC, which is a real challenge for me when I'm looking at loop rates and latency.

Taro: I'm interested in how they handle the uncertainty aspect; if we're moving away from first-principles models, the robustness guarantees are usually where things get tricky for autonomy.

Rosa: Exactly, and this paper proposes a way to build robust control even when we don't have perfect mathematical descriptions of our systems.

Dev: The paper outlines three specific data-driven ways to create these approximate DC models, which is interesting because it moves the modeling burden from purely theoretical methods onto learning techniques.

Taro: Learning the dynamics in a DC form sounds promising for real-world deployment because it means we can use actual sensor data to inform the optimization structure, rather than just guessing a model.

Rosa: They show methods like fitting polynomials to data, using input-convex neural networks, and even employing radial basis functions with specific kernel properties to achieve this DC representation.

Dev: The ICNN approach specifically mentions constraining the kernel weights to be non-negative during training and using convex activation functions, which is a neat way to keep the structure convex from the start.

Taro: If they can successfully derive these models online or near-online, it opens up possibilities for AI systems in areas where high-fidelity simulators are too slow to run continuously.

Rosa: That's what excites me about it; imagine using this for something like a PVTOL aircraft, which is inherently nonlinear and has continuous dynamics.

Dev: The core of the scheme involves a tube-based MPC algorithm that convexifies the online optimization by linearizing only the concave components of the model, which should keep the problem tractable at each time step.

Title and authors: Taro: I want to know how this works when things go wrong in an unpredictable environment; does it still maintain stability when we hit something unexpected?

Rosa: They provide rigorous guarantees for recursive feasibility and robust stability, which is a big deal because standard MPC often struggles with those guarantees under disturbance.

Dev: The framework relaxes the non-convex dynamic constraint into a convex form using the DC decomposition, specifically by linearizing the concave part of that decomposition around a predicted trajectory.

Taro: That linearization step sounds crucial; it means the system is guaranteed to behave well locally around where it expects to be, which is vital for safety when world conditions change.

Rosa: And they also address additive disturbances by modifying the algorithm with a backtracking line search scheme, ensuring recursive feasibility even when external noise messes things up.

Dev: The stability under those additive disturbances is guaranteed by Theorem seven which shows that the average stage cost stays bounded according to a specific quadratic stability condition involving t to infinity one over t X t-one n=zero xn - x r 2Q + un - u r 2R beta.

Taro: Bounded average stage cost is good, but I'm thinking about the limits of this robustness; does this framework still hold up if the modeling error in our data-driven DC model gets too large?

Rosa: The paper does state that the method provides guarantees for recursive feasibility and robust stability, but they also noted that first-principles models aren't available in DC form except in special cases, so the success really depends on how well their chosen data-driven approach captures the true dynamics.

Dev: They did compare performance using a planar vertical take-off and landing PVTOL aircraft case study, which gives us some concrete numbers to judge how much computational saving they actually achieve over traditional solvers.

Taro: If this works outside the lab for extended periods, that would be fantastic for autonomous systems operating in remote areas where we can't constantly re-tune parameters.

Rosa: That's the main question for me; right now, I see it primarily as a framework to solve complex problems offline or in highly controlled environments first.

Title and authors: Dev: The computational efficiency aspect is also highlighted, particularly when they use simplex parameterizations for the tube cross-section, which makes the optimization problems scale linearly with the number of states instead of exponentially.

Taro: That linear scaling is what makes it viable for resource-constrained hardware; if we can solve this in real time on an embedded system, that changes how quickly we can react to dynamic hazards.

Rosa: It seems like they've made a strong case for using DC programming as a way to bridge the gap between the high accuracy of nonlinear dynamics and the tractability needed for real-time control.

Dev: Before we wrap up, I want to make sure everyone has weighed in on what these results mean for practical implementation versus theoretical guarantees.

Taro: I'm just thinking that if we can reliably model complex dynamics this way, it means autonomy becomes much more dependable when facing unexpected world events.

Rosa: I agree with Taro; the dependability part is where this research truly has its potential impact on physical robotics and autonomous vehicles.

Dev: So, to wrap up on "Computationally Tractable Robust Nonlinear Model Predictive Control using DC Programming," they've developed a method to represent complex nonlinear systems in a difference-of-convex form using data-driven techniques, leading to a tube MPC scheme that guarantees recursive feasibility and robust stability even with additive disturbances.

Taro: That sounds like it could significantly advance the ability of autonomous agents to operate safely in complex, dynamic real-world settings by providing mathematically sound control guarantees.

Rosa: I think the three data-driven modeling procedures they presented—polynomial fitting, ICNNs, and RBFs—are really the most important parts for anyone looking to use this framework practically.

Dev: Precisely; we need those specific methods to choose based on whether our system dynamics are better represented by a polynomial approximation or a neural network structure.

Taro: And as long as these methods can accurately capture the system's nonlinearities, the impact could be felt across many domains, not just in one specific type of robot control.

Rosa: Indeed; this paper lays down a solid foundation for using learned models to create robust, real-time control policies that are much more flexible than those based on fixed physical equations alone.

The paper's summary: Rosa: So, to recap, this paper is about using difference-of-convex programming to make robust nonlinear model predictive control computationally feasible by leveraging data-driven methods for modeling those dynamics.

Dev: Right, that’s the big idea—taking a problem that usually explodes in complexity and transforming it into something solvable with convex optimization techniques.

Rosa: It sounds like they tackled the core issue of getting robustness guarantees in nonlinear MPC without making the computation time completely unusable for real-time applications.

Dev: Exactly, and I'm really interested in how they manage that trade-off between the quality of robustness and the required loop rate; that’s where I usually get stuck with traditional solvers.

Taro: From my angle, if we can actually guarantee stability even when things go wrong in the environment, that opens up a lot of possibilities for autonomous systems operating in unpredictable settings.

Rosa: It's exciting because they show concrete methods—like fitting polynomials to data or using input-convex neural networks—to build these DC models from scratch rather than relying on perfect physical equations.

Dev: And that’s where the engineering challenge comes in; we need to ensure those learned models don't introduce too much error that invalidates the robustness guarantees they claim.

Taro: I’m curious about how resilient this whole setup is when we introduce unexpected disturbances, which is always a huge factor in real-world autonomy.

Rosa: They even extended the algorithm to handle additive noise by incorporating a backtracking line search, which helps maintain recursive feasibility even when the predicted trajectory gets hit by external forces.

Dev: That's a smart move; ensuring that the system doesn't just fail because of minor noise is crucial for any control engineer evaluating this.

Taro: If this framework can reliably handle those kinds of disturbances, imagine how much more dependable our autonomous agents become when they’re navigating cluttered or dynamic physical spaces.

Rosa: That’s what I think about; it moves us closer to having robotic systems that don't just work perfectly in a controlled lab but can operate safely out there for longer periods.

Dev: And the computational efficiency they achieve by exploiting the DC decomposition, especially when using simplex parameterizations for those tube sets, seems like a really practical win for deployment on embedded hardware.

Taro: So if these learned models can provide that level of safety and tractability, what does this imply for the next generation of complex robotic tasks?

Rosa: It implies we can start designing controllers based on data-driven dynamics instead of waiting for perfect first-principles models to emerge, which should accelerate development across many domains.

Dev: I’m looking forward to seeing the specifics in the experiments; it's one thing to have a theoretical guarantee, but proving it holds up under stress is what matters for me.

Taro: Let’s see if they can push this framework beyond just stability and into more complex, goal-oriented planning scenarios.

The paper's improvements: Rosa: So, to wrap up on the improvements, it seems the paper isn't just proposing one method but actually offering three distinct data-driven approaches for creating those tractable DC models: fitting polynomials, using input-convex neural networks, and employing radial basis functions.

Dev: That variety is interesting because it means the framework can be applied to different types of nonlinear system dynamics depending on what kind of data you have available for training.

Rosa: Right, and they also showed how to convexify the online optimization by specifically linearizing only the concave parts of the model constraints, which keeps things efficient without sacrificing too much accuracy.

Dev: I like that approach; it suggests we don't need to linearize the whole thing around every single point, just where it’s mathematically convenient, which helps keep those loop rates manageable.

Taro: If this works well in theory, the implication is that we can move towards building control policies for systems where we don't have a complete mathematical description yet.

Rosa: Exactly; it means AI can start learning how to control physical processes directly from raw data rather than waiting for highly idealized simulations.

Dev: I’m thinking about the stability guarantees they provide, especially under those additive disturbances that we talked about earlier; those robustness proofs are pretty solid if they hold up in practice.

Taro: If the system can handle external noise reliably, it would be a big step for autonomy because real-world environments are messy and unpredictable.

Rosa: I agree with Taro; the fact that they’ve shown how to guarantee recursive feasibility even when things go wrong makes this framework much more appealing for deployment outside of controlled lab settings.

Dev: The computational efficiency gains, especially when using simplex parameterizations for the tube cross-sections, are huge; that means lower latency, which is essential for safety-critical control.

Rosa: So we’re talking about a system that can be learned from data and then optimized in real time with high reliability, which feels like a significant step forward for field robotics.

Taro: I wonder how long these models can maintain their performance if the underlying physical system slowly degrades over time due to wear or aging.

Dev: That’s a limitation they actually flag; the paper focuses on the robustness of the *control scheme* given an initial model, but it doesn't inherently solve continuous model drift without further adaptation mechanisms.

Rosa: It sounds like future work will need to focus on integrating those adaptive mechanisms more tightly into the DC formulation itself so that these models stay accurate over long operational periods.

Conclusion: Rosa: To wrap up, this paper on "Computationally Tractable Robust Nonlinear Model Predictive Control using DC Programming" shows how we can use data-driven modeling to make nonlinear control problems manageable for real-time applications by transforming them into tractable difference-of-convex forms.

Dev: That’s a solid summary; it highlights the shift from intractable nonconvex optimization to methods that allow for guaranteed stability under uncertainty, which is exactly what we need for reliable loop rates.

Rosa: It really opens up avenues for field robotics, I think; if these learned models can operate reliably outside the lab, that’s where we need to be seeing this kind of work applied.

Dev: I'm still focused on how they manage the actual latency during those sequential convex steps; that part needs serious engineering validation before we can put it on a production system.

Taro: From an autonomy standpoint, the fact that this framework addresses robustness against external disturbances is very encouraging because it suggests agents can maintain their intended behavior even when things go unexpectedly.

Rosa: I agree with Taro; having a control scheme that maintains recursive feasibility when the world misbehaves is a huge deal for any autonomous system we design.

Dev: The efficiency gains they show, particularly with simplex parameterizations, mean we can run these complex checks much faster than traditional solvers allow for high-frequency updates.

Taro: If this translates well into more complex scenarios, it could significantly enhance the reliability of systems dealing with dynamic environments or cluttered spaces.

Rosa: Overall, I think the biggest implication is that we’re getting closer to building control policies that are both highly accurate and computationally lean enough for actual deployment.

Dev: I’m optimistic about this direction; seeing these structured convex programs solves a lot of the failure mode issues we usually encounter when dealing with high-dimensional nonlinear systems.

Taro: I'm just thinking about the long-term vision; if we can use learned DC models for control, it could pave the way for more generalized autonomous agents that aren't tied to specific physical system parameters.

Martin Doff-Sotta, Zaheen A-Rahman, Mark Cannon

Department of Engineering Science, University of Oxford

math.OC, cs.SY, eess.SY

Submitted: 2026-02-01

Updated: 2026-09-29

Code: https://github.com/martindoff/DC-NN-MPC

License: http://creativecommons.org/licenses/by-nc-nd/4.0/

Importance score: 79/100

The gist: This paper proposes a computationally tractable robust Model Predictive Control (MPC) framework for nonlinear systems by leveraging difference-of-convex (DC) programming and sequential convex

Key concepts

Difference of Convex (DC) Programming: This is a mathematical technique used to solve optimization problems where the objective function or constraints are not entirely convex, but can be expressed as the difference between two convex functions. The paper uses this to transform a difficult nonlinear problem into one that can be solved efficiently using standard convex optimization methods.
Data-Driven DC Modelling: This refers to creating an approximate mathematical model of a real-world nonlinear system by fitting data. Three methods are used: fitting polynomials, using input-convex neural networks, or employing radial basis functions. These methods aim to represent the complex system dynamics as a DC function for easier online optimization.
Tube MPC Algorithm: This is the control strategy where the controller calculates a sequence of inputs that keeps the system's state within a 'tube' around a desired trajectory. The tube's shape is defined by robustly invariant sets, which helps ensure stability even when disturbances occur, by linearizing concave parts of the model.

Terminology

Model Predictive Control — 17× in this paper
Robustness — 12× in this paper · explained in 5 episodes
Input-Convex Neural Network — 10× in this paper
Accuracy — 5× in this paper · explained in 1 episode
Activation Function — 5× in this paper · explained in 1 episode
Core — 3× in this paper · explained in 2 episodes
Model Predictive Control (MPC) — 3× in this paper · explained in 5 episodes
Semidefinite Program (SDP) — 3× in this paper · explained in 2 episodes
Computational overhead — 2× in this paper · explained in 2 episodes
Adaptation — 1× in this paper · explained in 1 episode
Adaptive Mechanism — 1× in this paper · explained in 1 episode
Fidelity — 1× in this paper · explained in 2 episodes
Lyapunov function — 1× in this paper · explained in 2 episodes
Model Drift — 1× in this paper · explained in 4 episodes
Objective Function — 1× in this paper · explained in 2 episodes
Theoretical Guarantees — 1× in this paper · explained in 2 episodes

Summary

This paper proposes a computationally tractable robust Model Predictive Control (MPC) framework for nonlinear systems by leveraging difference-of-convex (DC) programming and sequential convex programming. It addresses the challenge of achieving robustness guarantees in nonlinear MPC, which are often hindered by high computational costs or overly conservative designs. The authors develop a data-driven modeling approach to represent system dynamics in DC form, allowing the online optimization problem to be solved using convex methods, thereby improving the trade-off between computational cost and robustness for systems with differentiable discrete-time dynamics.

Data-Driven DC Modelling Approaches

The paper outlines three distinct data-driven procedures for constructing approximate DC models of nonlinear systems. These methods aim to express a continuous function as a difference of two convex functions, which is essential for formulating the online MPC optimization problem as a difference of convex (DC) programming problem.

Fit polynomial to data: This method involves approximating the nonlinear model with a polynomial of degree 2d in Gram form, fitting it to data samples via least squares, and then solving a Semidefinite Program (SDP) to find convex polynomials for the functions g and h such that f = g - h.
Input-convex neural networks: This approach utilizes an Input-Convex Neural Network (ICNN), where kernel weights are constrained to be non-negative at training, and convex activation functions are used. A DCNN structure, consisting of two parallel ICNNs with subtracted outputs, is proposed to learn the dynamics in DC form.
Radial basis functions: This method approximates the function using a weighted sum of Radial Basis Functions (RBFs). If the RBF kernel is convex (i.e., its second derivative and first derivative are non-negative), the function can be expressed as a DC function: f = g - h, where g and h are defined using max and min operations involving the RBF kernel.

Tube MPC Algorithm Development

The core of the proposed control scheme is a tube-based MPC algorithm that convexifies the online optimization by linearizing concave components of the model. The scheme results in a sequence of computationally tractable convex programs to be solved at each time step, based on successive linearization and exploiting the convex structure in the dynamics.

The general TMPC paradigm considers a two degree of freedom control law:

(3) u k = v k + K k x k

where the feedforward control sequence stabilizes uncertain state trajectories within a tube whose cross-sections are robustly invariant sets parameterized by polytopic sets. The tubes can be parameterized using either elementwise bounds or a simplex parameterization, resulting in different numbers of vertices for the tube cross-section.

Optimization Formulation and Convexification

The non-convex dynamic constraint (6) is relaxed into a convex form (7) by exploiting the DC decomposition:

(7) g(x k, v k + K k x k) - ⌊h◦⌋(x k, v k + K k x k) ≤ qn+1

where ⌊h◦⌋ is the Jacobian linearization of the concave part of the DC decomposition around a predicted trajectory. This convex inequality is then combined with state and input constraints (8) to form a convex program (16). The worst-case cost function, which involves maximum operations over the tube sets, is reformulated using slack variables θ, χ, and by evaluating it at the vertices of the tube cross-section.

Robustness Guarantees and Stability Results

The framework provides rigorous guarantees for recursive feasibility and robust stability.

(Proposition 2 (Feasibility))

If an initial feasible trajectory exists at time n=0, problem (16) is feasible at all times n > 0 and all iterations j > 0.

(Theorem 5 (Asymptotic stability))

For a system in DC form controlled by Algorithm 1, the equilibrium x = x r is an asymptotically stable equilibrium with a region of attraction equal to the set of initial states for which a feasible initial trajectory exists.

When additive disturbances are present, the algorithm is modified using a backtracking line search scheme (Algorithm 2) to ensure recursive feasibility. The stability under additive disturbances is guaranteed by Theorem 7, which shows that the average stage cost remains bounded:

(Theorem 7 (Stability with disturbance))

The control law ensures that the system satisfies state and input constraints and the following quadratic stability condition on the average stage cost holds: limt→∞ 1/t Xt−1 n=0 x[n] − x r2Q + u[n] − u r2R ≤ β, where β satisfies equation (21).

Improvements for AI systems

Based on the provided scientific paper, here are specific improvements that can be made to AI systems by implementing its proposed framework, and what these improved systems could achieve:

Improved Robust Control for Nonlinear Systems: The core improvement is the development of a computationally tractable Robust Model Predictive Control (MPC) algorithm that handles nonlinear dynamics with uncertainty.
System Capability: This system can now perform real-time, safe control of complex nonlinear physical systems (like PVTOL aircraft or other continuous-dynamics systems) while explicitly guaranteeing stability and constraint satisfaction even when the underlying model is inaccurate or external disturbances are present.
Enhanced Model Representation via Data-Driven DC Functions: The framework provides three specific, data-driven methods to convert complex, unknown nonlinear dynamics into a structure amenable to optimization (Difference of Convex, or DC form):
System Capability: AI systems can now learn and represent the dynamics of physical processes (e.g., fluid dynamics, chemical reactions) directly from experimental data using techniques like Sum-of-Squares (SOS) polynomials, Input-Convex Neural Networks (ICNNs), or Radial Basis Functions (RBFs). This allows for the creation of highly accurate, yet mathematically tractable models that are superior to traditional first-principles models when they are unavailable or contain unknown parameters.
Efficient and Guaranteed Online Optimization: The paper introduces a method to convexify the online MPC optimization by exploiting the convex-concave procedure (CCP), which avoids conservative linearization error bounds by only linearizing concave components of the model constraints.
System Capability: This allows AI controllers to solve complex, high-dimensional, nonlinear optimization problems in real-time with guaranteed convergence properties (recursive feasibility and robust stability). This is crucial for deployment in safety-critical systems where traditional nonconvex solvers are too slow or risk instability.
Adaptive Disturbance Rejection (Additive Uncertainty): The algorithm is extended to handle additive disturbances by incorporating a heuristic based on a backtracking line search to ensure recursive feasibility, even when the predicted trajectory becomes infeasible due to external noise or modeling errors.
System Capability: AI systems can maintain robust performance in the presence of real-world, unpredictable noise and modeling inaccuracies. The system does not crash or become unstable if it encounters unexpected external forces, as it can dynamically adjust its control sequence (via the backtracking search) to remain feasible within a defined robust tube.
Computational Efficiency for Real-Time Use: The paper demonstrates that using simplex parameterizations for the tube cross-section results in optimization problems with only linearly scaling inequalities with the number of states, making them computationally efficient compared to elementwise bounds (which scale exponentially). Furthermore, the scheme allows for early termination after just one iteration per time step without sacrificing stability.
Optimized Control Law Computation: The paper provides explicit formulas (via Dynamic Programming recursion) for computing the necessary feedback gains and terminal parameters using Semidefinite Programming (SDP).
System Capability: This enables the deployment of highly scalable, low-latency control systems in resource-constrained environments (like autonomous vehicles or embedded hardware) because the computational burden is predictable and minimized through structured convex programming.

In summary, the improved AI system can be a high-performance, data-driven control agent capable of operating safely and reliably in complex, uncertain physical environments by leveraging learned dynamics expressed in a mathematically tractable DC form.

Abstract

We propose a computationally tractable, tube-based robust nonlinear model predictive control (MPC) framework using difference-of-convex (DC) functions and sequential convex programming. For systems with differentiable discrete time dynamics, we show how to construct systematic, data-driven DC model representations using polynomials and machine learning techniques. We develop a robust tube MPC scheme that convexifies the online optimization by linearizing the concave components of the model, and we provide guarantees of recursive feasibility and robust stability. We present three data-driven procedures for computing DC models and compare performance using a planar vertical take-off and landing (PVTOL) aircraft case study.