An Integrated Architecture for Data-Efficient Real-Time Neural Imitation of Nonlinear Time-Optimal Model Predictive Control
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.
Rosa: We're looking at the paper: "An Integrated Architecture for Data-Efficient Real-Time Neural Imitation of Nonlinear Time-Optimal Model Predictive Control".
Dev: Here's the gist of the paper. Nonlinear time-optimal model predictive control (TOMPC) enables highly performant control of constrained systems, but its computational cost prohibits real-time use on fast dynamical systems.
Rosa: This work imitates a long-horizon nonconvex TOMPC with hard terminal constraints whose high computational cost limits training-data generation.
Dev: The authors present a complete training and control architecture that jointly addresses dataset correlation, suboptimal expert labels, covariate shift, constraint satisfaction and model mismatch.
Rosa: They introduce self-informed sampling (SIFS), a data-efficient expert-sampling strategy that uses a learned time-to-go to query the expert only where the policy deviates from time-optimal behavior.
Paper summary: Rosa: So to wrap up the summary of "An Integrated Architecture for Data-Efficient Real-Time Neural Imitation of Nonlinear Time-Optimal Model Predictive Control," the paper tackles the problem that nonlinear time-optimal model predictive control is too computationally expensive for real time use, even though it gives great performance on constrained systems.
Dev: The central thesis is that imitating this hard problem, which has high computational cost and limits training data generation, requires a complete architecture to handle all the usual headaches: data correlation, suboptimal expert labels from nonconvex problems, covariate shift in the model itself, constraint satisfaction issues, and plant model mismatch.
Taro: They propose three main things to solve this: first self-informed sampling which uses a learned time-to-go to query the expert only when the policy is not following time optimal behavior.
Rosa: Second, they present this complete training and control architecture that manages all those challenges jointly while also improving how they sample the expert data.
Dev: And third, they show real-world validation on a physical Furuta pendulum where their neural network imitation controller actually swings the system up and stabilizes it along trajectories matching simulation results.
Taro: This work essentially shows how you can use AI to create a fast, reliable controller for hard problems that were previously computationally intractable because you solved the sampling efficiency, labeling conflicts, and control robustness issues all together.
Conclusion: Rosa: Thinking about the title of "An Integrated Architecture for Data-Efficient Real-Time Neural Imitation of Nonlinear Time-Optimal Model Predictive Control," it really sounds like they’re describing a unified system.
Dev: Right. The implication here is that we can build systems that have high performance and tight constraints, but instead of needing a massive amount of expert data to train the model, we can use smarter sampling and better training techniques to achieve similar results with less data.
Taro: It means for applications requiring real-time control on fast dynamics, we might be able to use these powerful neural network imitators much sooner than before because they tackle the data generation bottleneck directly.
Rosa: So in short, this paper gives us a blueprint for building robust, fast controllers by integrating data efficiency with a complete framework that handles the complexities of real-world control problems.
Maarten De Baecke, Flavia Sofia Acerbo, Andrea Giusti, Wilm Decreä
Department of Mechanical Engineering at KU Leuven · Flanders Make vzw
eess.SY, cs.LG, cs.RO, cs.SY
Submitted: 2026-11-12
Updated: 2026-11-12
Comments: Submitted to European Control Conference 2027 (Brussels)
Project page: https://maartendebaecke.github.io/PW/neural-tompc
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 83/100
The gist: Nonlinear time-optimal model predictive control (TOMPC) enables highly performant control of constrained systems, but its computational cost prohibits real-time use on fast dynamical systems.
Key concepts
- Nonlinear Time-Optimal MPC (TOMPC)
- This is an advanced control method designed to find the best sequence of control actions over a long period, aiming to minimize time. It uses a complex mathematical formulation with hard terminal constraints, which makes it highly accurate but computationally very expensive for fast systems.
- Self-Informed Sampling (SIFS)
- This is a data-efficient strategy where the neural network learns to query the expert only when its current control policy deviates from the ideal time-optimal behavior. It uses a learned 'time-to-go' value function to intelligently select which expert data points are most valuable for training.
- Safety Filter
- This is a final component in the neural network that guarantees safety by projecting the raw network output onto a set of inputs known to keep the physical system within its required constraints at all future times. It works alongside a penalty term in the training loss to ensure feasibility.
- Ancillary Controller
- This is a lightweight controller added to the neural network output. Its job is to compensate for inaccuracies between the model used in the neural network and how the actual physical system behaves, helping the controller generalize better when faced with real-world plant mismatches.
Terminology
Summary
Nonlinear time-optimal model predictive control (TOMPC) enables highly performant control of constrained systems, but its computational cost prohibits real-time use on fast dynamical systems. p. 0 This work imitates a long-horizon nonconvex TOMPC with hard terminal constraints whose high computational cost limits training-data generation. p. 0 The authors present a complete training and control architecture that jointly addresses dataset correlation, suboptimal expert labels, covariate shift, constraint satisfaction and model mismatch. p. 0 They introduce self-informed sampling (SIFS), a data-efficient expert-sampling strategy that uses a learned time-to-go to query the expert only where the policy deviates from time-optimal behavior. p. 0 The resulting controller matches the region of attraction of the expert TOMPC while raising the control frequency from 1. p. 04 Hz to 1 kHz. p. 0
Expert Model Predictive Controller
The expert MPC considered here involves a long-horizon, nonconvex, time-optimal formulation with a hard terminal constraint. Time-optimality is realized using an exponentially weighted objective following [2]. p. 2 For TOMPC, the time-to-go (T2G) naturally serves as a value function defined as the first time the predicted MPC trajectory from an initial state x0 lies within a tolerance ξ of the reference xf: T2G(x0) = ∆t ·min[k: ∥xk −xf∥ ≤ ξ]. p. 2
How the training signal is built
The proposed framework builds upon the approach of [12] in which the NN is trained on an Optimally-Sampled Dataset preserving the information content of the MPC policy at a prescribed resolution while eliminating redundant data points. The training loop involves generating new data points from an expert MPC I, iteratively aggregating them to dataset II, filtering with an optimal sampling algorithm III to remove correlated and redundant samples, assigning weights IV to reduce the influence of conflicting expert labels, and then training the NN V on the weighted dataset using BC. p. 1
Where the gains come from
The key contributions include three original elements tailored to TOMPC:
-
Self-informed sampling (SIFS), a data-efficient expert-sampling strategy that uses a learned time-to-go to query the expert only where the policy deviates from time-optimal behavior. p. 0
-
A complete training and control architecture that jointly addresses dataset correlation, suboptimal expert labels, controller adaptability, constraint satisfaction and covariate shift and efficient expert sampling. p. 0
-
Real-world validation on a physical Furuta pendulum. p. 0
Component Validation
The NNMPC output is adjusted by an ancillary controller VIII for plant–model mismatch compensation and a safety filter IX for constraint satisfaction. The training loss is formulated as Ltotal = λuLu +λT2GLT2G +λbLbarrier +λpLpenalty. p. 4 The safety filter projects the raw network output onto the set of inputs that keep the system within its constraints at all future times. p. 4
Hardware Validation
On hardware, the controller swings up and stabilizes the pendulum along trajectories closely matching simulation. The NNMPC controller supports a maximum control frequency of ≈ 100 Hz on CPU and consistently remains below 1 ms inference time on GPU, theoretically enabling control frequencies up to 1 kHz. p. 5 The resulting NNMPC attains a region of attraction on par with the expert MPC at an over 700-fold lower worst-case computation time (on GPU). p. 6 p. 6 p. 6 The controller performance for the physical pendulum was assessed through a swing-up task where the real-world state trajectory closely matches the simulated results. p. 6 p. 6 p. 6
Conclusion
We presented an integrated architecture for imitating TOMPC with a neural network which combines state-of-the-art techniques for dataset correlation, constraint satisfaction, and covariate shift with three original contributions tailored to TOMPC. The self-informed sampling matched the performance of state-of-the-art methods with less than half the training data and roughly one third of the training data. p. 6 The resulting NNMPC attains a region of attraction on par with the expert MPC at an over 700-fold lower worst-case computation time (on GPU). The ancillary controllers substantially improved generalization to perturbed plant models in simulation. p. 6 The controller performance for the physical pendulum was assessed through a swing-up task where the real-world state trajectory closely matches the simulated results. The main discrepancy is a high-frequency oscillatory control effort near the inverted equilibrium attributable to sensor noise and residual plant–model mismatch. p. 6
Pruning redundant reasoning tokens
The paper details the challenges of imitating a long-horizon, time-optimal formulation with hard terminal constraints because its high computational cost limits training-data generation. Time-optimal control laws tend toward bang-bang behavior requiring dense data near switching points which increases the demand for expert data. p. 0 Existing works address individual challenges such as dataset correlation, suboptimal expert labels, controller adaptability, constraint satisfaction and covariate shift but lack a complete integrated architecture. p. 0
Suboptimal Data
To mitigate conflicting expert labels from nonconvex optimal control problems where nearby states can receive conflicting expert labels, this work employs weighted BC multiplying each quadratic error term in the training loss by a weight. This weighting strategy is computed using a lightweight T2G-based strategy assigning each training sample a weight wknn(xi) based on its T2G relative to the distance-weighted average T2G of its M nearest neighbors. p. 2
Constraint Satisfaction
Constraint satisfaction is addressed at two levels: additional terms in the training loss encourage the neural network to learn constraints imposed by the original MPC formulation, while a differentiable safety filter in the final network layer formally guarantees constraint satisfaction. The training loss combines a logarithmic barrier with a quadratic penalty to steer raw network output towards the feasible region. p. 3
Self-Informed Sampling
We introduce self-informed sampling as a data-efficient sampling strategy that addresses the limitations of existing approaches by augmenting the NN with an additional output head that predicts a value function, more specifically the T2G. The informativestate criterion should therefore tolerate short-term deviations and isolated jumps as long as the overall decreasing trend remains consistent. p. 2 Candidates are then sampled for expert querying with a probability that increases with their proximity to a flag, how early they occur in the rollout, and the severity of the corresponding rollout. p. 3
Controller Adaptability
To compensate for plant–model mismatch without retraining on real-world data, the NNMPC is augmented with a lightweight ancillary controller that corrects its control action online. Two options are considered: a dynamic-inversion compensator (DIC) and a Linear Quadratic Regulator (LQR). p. 3 The DIC analytically inverts the plant dynamics to find the input u¯ that reproduces the desired output derivative on the real plant. p. 3 The LQR directly converts this observed mismatch into a corrective input using a computed LQR gain K. p. 3
Final NNMPC
The final NNMPC output uk safe is obtained by passing the nominal control output through the safety layer to obtain the applied control input uk safe. The multi-objective loss is formulated as Ltotal = λuLu +λT2GLT2G +λbLbarrier +λpLpenalty. p. 4 The training procedure involved pre-training on an initial grid-based dataset followed by iterative training with optimal sampling and weighting. p. 4 The controller performance for the physical pendulum was assessed through a swing-up task where the real-world state trajectory closely matches the simulated results. The resulting NNMPC attains a region of attraction on par with the expert MPC at an over 700-fold lower worst-case computation time (on GPU). The controller swings up and stabilizes the pendulum along trajectories closely matching simulation. p. 6
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Data-efficient expert sampling using self-informed sampling (SIFS) improves data efficiency by
matching standard dataset aggregation performance in roughly one-third of the training time and with less than half the training data.
This allows for faster model training while maintaining high performance compared to traditional methods like DAgger or Monte Carlo dropout. -
Joint architecture integration addresses multiple challenges simultaneously, including
dataset correlation, suboptimal expert labels, covariate shift, constraint satisfaction and model mismatch.
This holistic approach leads to a more reliable Neural Network Model Predictive Controller (NNMPC) system by ensuring robustness across various learning difficulties. -
The controller adaptability component uses an ancillary controller like the Dynamic-Inversion Compensator (DIC) or Linear Quadratic Regulator (LQR)
to compensate for plant–model mismatch without retraining on real-world data.
This enables the NNMPC to generalize better to different physical systems by correcting errors online using analytical inversion or linear feedback. -
Safety guarantees are achieved through a combination of a differentiable safety filter and training loss terms, specifically combining a logarithmic barrier with a quadratic penalty. This ensures constraint satisfaction by providing
a gradient in the infeasible region, thereby mitigating this instability
during training and guaranteeing feasibility during runtime via the safety filter. -
The improved NNMPC system can operate at significantly higher control frequencies, achieving inference times below 1 millisecond on GPU, theoretically enabling
control frequencies up to 1 kHz.
This capability allows for real-time deployment on fast dynamical systems where the original MPC's frequency of approximately 1.36 Hz is insufficient.
Sources
- SCIZOR: A Self-Supervised Approach to Data Curation for Large-Scale Imitation Learning
- Transfer from Simulation to Real World through Learning Deep Inverse Dynamics Model
Related papers
- Industrial Load Modeling and Optimization for Market-Based Interaction with Power Systems
- Exact and Approximate Convex Reformulation of Linear Stochastic Optimal Control with Chance Constraints
- Certified Set Convergence for Piecewise Affine Systems
- Symbolic Control for Stochastic Systems via Finite Parity Games
- Multiobjective optimization-based design and dispatch of islanded, hybrid microgrids for remote, off-grid communities in sub-Saharan Africa
- Physiology-informed layered sensing for intelligent human-exoskeleton interaction