An Integrated Architecture for Data-Efficient Real-Time Neural Imitation of Nonlinear Time-Optimal Model Predictive Control
- 1KU Leuven, Dept. of Mechanical Engineering, MECO Research Team & Flanders Make@KU Leuven
- 2Flanders Make vzw, Lommel, Belgium
Citation show BibTeXhide
@article{debaecke2026neural,
title = {An Integrated Architecture for Data-Efficient Real-Time Neural
Imitation of Nonlinear Time-Optimal Model Predictive Control},
author = {De Baecke, Maarten and Acerbo, Flavia Sofia and
Giusti, Andrea and Decr{\'e}, Wilm},
journal = {TBD},
year = {2026},
note = {In preparation}
}
Placeholder. The final reference will appear here once the paper is published.
Abstract readhide
Nonlinear time-optimal model predictive control (MPC) enables highly performant control of constrained systems, but its computational cost prohibits real-time use on fast dynamical systems. Neural network imitation of the MPC control law can reduce online computational costs, yet existing work mostly targets quadratic-tracking formulations in simulation and addresses the challenges of such controllers in isolation. This work imitates a long-horizon, nonconvex, time-optimal MPC with hard terminal constraints, whose high computational cost strongly limits training-data generation. We present (i) 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, (ii) a complete training and control architecture that jointly addresses dataset correlation, suboptimal expert labels, covariate shift, constraint satisfaction and model mismatch, and (iii) real-world validation on a Furuta pendulum. SIFS matches the performance of standard dataset aggregation about three times faster and with less than half the data. The resulting controller attains a region of attraction on par with the expert MPC while raising the reliable control frequency from 1.4 Hz to 1 kHz. On hardware, the controller swings up and stabilizes the pendulum along trajectories closely matching the expert MPC solution.
Hardware experiments
- Reference
- pendulum upright, base angle zero
- Constraint
- −π/2 ≤ base angle ≤ π/2
Contributions
- Self-informed sampling (SIFS). A data-efficient alternative to DAgger. The network predicts its own time-to-go, and a break in the expected linear decrease flags the states where the policy deviates from time-optimal behavior. Only those states are sent to the expert.
- A complete training and control architecture. One pipeline that jointly handles dataset correlation, conflicting expert labels, covariate shift, constraint satisfaction and plant–model mismatch, problems that earlier work tackled one at a time.
- Real-world validation. Deployment on a physical Furuta pendulum, imitating a long-horizon, nonconvex time-optimal MPC with hard terminal constraints.