An Integrated Architecture for Data-Efficient Real-Time Neural Imitation of Nonlinear Time-Optimal Model Predictive Control
TL;DR Imitate a slow time-optimal MPC for swinging up a Furuta pendulum with a small neural network, querying the expert only where the network's own predicted time-to-go stops falling as it should. This matches DAgger with less than half the data in about a third of the training time, and the network runs at 250 Hz on hardware where the MPC manages 1.36 Hz.
What they do
A time-optimal MPC can swing a Furuta pendulum up and stabilize it within one 60-step horizon, but a solve takes 114 ms on average and up to 733 ms. The authors imitate it with a small neural network. Expert solves are expensive and the bang-bang control law needs dense data near its switching points, so the key question is where to query the expert; the paper also combines existing fixes for imitating MPC into one architecture.
How it works
- Self-informed sampling (SIFS). A second output head predicts time-to-go (T2G), which should fall by one sampling interval per step along a time-optimal rollout. Where the prediction stagnates, rises or jumps, states at and just before that point are sent to the expert.
- Dataset handling. Correlated samples are thinned by optimal sampling. Conflicting labels from different local optima are down-weighted by comparing each sample’s T2G with that of nearby states from other trajectories.
- Deployment. A barrier-plus-penalty loss and a projection safety filter keep inputs within constraints. An added dynamic-inversion compensator or LQR corrects plant-model mismatch without retraining.
Results
- Over 1000 random initial states in simulation, both controllers converge in 91.5% of cases, only the network in 4.2% and only the MPC in 4.3%. The median time to the target is 1.62 s, against 1.35 s for the MPC.
- SIFS reaches the failure rate of DAgger and MC-dropout sampling about three times faster and with less than half the data. Removing the T2G weighting raises the failure rate.
- Network inference takes 0.41 ms on GPU (1.0 ms worst case). On the physical pendulum, at 250 Hz, the swing-up closely follows simulation.
Limitations
All experiments use one system, the Furuta pendulum, and the mismatch compensators are tested only in simulation. The network still converges slightly slower than the expert, possibly from too little training or conflicting labels; labels from equally time-optimal modes are not resolved. On hardware, control effort oscillates near the upright position, which the authors attribute to sensor noise and model mismatch.
Videos
Figures



Abstract
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. Neural network imitation of the TOMPC control law can reduce online computational costs, yet existing work mostly targets quadratic cost formulations in simulation and addresses the challenges of neural controllers in isolation. This work imitates a long-horizon nonconvex TOMPC with hard terminal constraints, whose high computational cost limits training-data generation. We present (i) \emph{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 standard dataset aggregation performance in roughly one-third of the training time and with less than half the training data. The resulting controller matches the region of attraction of the expert TOMPC while raising the 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 TOMPC solution.
Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.
