Systems and Control · arXiv:2611.17005 v1 · submitted 12 Nov 2026 · 35 references
An Integrated Architecture for Data-Efficient Real-Time Neural Imitation of Nonlinear Time-Optimal Model Predictive Control
The paper says
“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.”
more of the abstract
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) 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.
from the abstract
Pith says
Machine review · deepseek-v4-flash · 13 Nov 2026 · what was read was not recorded · not checked by a person
It can be wrong. Challenge this review or reply to it.
Holds up, with conditions. SIFS is a genuinely new sampling idea for NNMPC, but the central data-efficiency claim is undercut by a one-sided stopping rule in the baseline comparison. This line is wrong
4 major objections · 8 minor
How often objections from reviews like this one are right: not yet measured
Not checked: figures, code, what the 35 references say.
The paper says, at Sec. VI.B (Fig. 5b); abstract; Sec. VIII
The review did not quote this passage.
Pith says: Objection 1 of 4 · major
The headline data-efficiency claim rests on a one-sided comparison. The caption of Fig. 5b states that the reported queries and training times are those 'needed to match the SIFS failure rate', so DAgger and MC-dropout are stopped at the first point at which they reach a failure rate that SIFS has already achieved; the paper does not report where those …
more of objection 1
… baselines end up if trained to the full 1441-iteration budget of Section IV.C, and it does not identify the SIFS checkpoint (iteration, failure-rate value, query count) used as the match target. Under this stopping rule, the abstract's claim that SIFS 'matches standard dataset aggregation performance' cannot be distinguished from the alternative that SIFS reaches an intermediate failure rate faster while converging to a worse final policy. The comparison also needs internal reconciliation: the full SIFS run aggregates 1,189,320 expert-labeled points over 1441 iterations, while Fig. 5b's query axis tops out at 9,000 additional queries, so the match checkpoint and its failure-rate value must be stated explicitly. Please run the baselines to the full budget, report their final failure rates and the failure rate at the matching SIFS checkpoint, and state the match criterion before the comparison is made.
The paper says, at Figs. 3-5; Sec. VII
The review did not quote this passage.
Pith says: Objection 2 of 4 · major
Figs. 4 and 5 report a single failure-rate curve per method with no variance across training seeds, and the region-of-attraction comparison in Fig. 3 (both controllers succeed in 91.5% of the 1000 sampled initial states) is a single-network point estimate; the hardware evidence in Section VII is five runs of a single swing-up task.
more of objection 2
For claims whose magnitude is expressed as factors (about one-third of the training time, less than half the data, and 91.5% parity), seed-to-seed variance is essential. Please report means and standard deviations (or min-max envelopes) over at least three to five seeds for the failure-rate curves, query counts, and times-to-match, and state whether the hardware reruns vary the initial condition or only the execution noise.
The paper says, at Sec. III.D; Sec. IV.A
The review did not quote this passage.
A classifier read this sentence of the paper as acknowledging the concern (probability 0.69):
“Since the T2G is learned, its predictions contain noise and occasional harmless discontinuities.”
Pith says: Objection 3 of 4 · major
The SIFS mechanism rests on the faithfulness of the learned T2G, and its justification at b100 is an informal plausibility claim. The argument that an exactly one-interval-per-step decrease is 'unlikely to occur by chance' is not quantified, and the paper's own acknowledgments of prediction noise and discontinuities (b88) and of expert suboptimality (b151) give concrete failure modes for the linear-trend …
more of objection 3
… template; the T2G labels are also capped by the N=60 horizon (1.8 s at dt=0.03), with no stated convention for the case where the expert's predicted trajectory does not enter the xi-ball within the horizon. Because the T2G is learned from expert labels and then used both to select query states and to weight conflicting labels, the loop is a bootstrap rather than a formal reduction; the central question is whether flags actually coincide with policy suboptimality. A concrete diagnostic would settle this: report the correlation between flagged states and one-step imitation error or closed-loop failure, or compare SIFS with a selection rule based on the ground-truth expert T2G residual, while holding the query budget fixed.
The paper says, at Sec. III.F; Secs. IV-V
The review did not quote this passage.
Pith says: Objection 4 of 4 · major
The formal constraint-satisfaction guarantee of Section III.F is conditional on a control-invariant set C that is never instantiated for the Furuta plant. The section acknowledges that the maximal control-invariant set may be intractable and offers CBFs as an alternative, but the paper does not state which option the experiments use, whether the projection onto U_s(x) is actually computed online at …
more of objection 4
… 250 Hz, or whether the base-angle and input box constraints were violated in the 1000-state evaluation of Fig. 3 or in the hardware runs. Consequently, the claim that the safety filter 'formally guarantees constraint satisfaction' is not verifiable for the demonstrated system. Please specify the concrete C (or the CBF and its relative-degree assumptions) used in the Furuta example, state how the projection's differentiability is handled in the training loss (9), and report constraint-violation statistics for the simulation and hardware evaluations.
8 minor objections
The paper says, at Sec. III.C, Eq. (2)
Pith says: minor objection 1 of 8
As printed, the exponent in Eq. (2) contains T2G(x_i) - T2G(x_i), which is identically zero, so the weight is constant and the equation cannot be the mechanism behind the ablation in Fig. 4; please correct the formula (presumably T2G(x_i) minus the neighbor-averaged T2G of Eq. (1)) and state the sign convention so that the most time-optimal samples receive the largest weight.
The paper says, at Fig. 5b
Pith says: minor objection 2 of 8
Please clarify what 'training time' includes (network retraining, expert queries, or both), whether the query counts are additional to the 15,601-point initial grid of Section IV.C, and what failure-rate value is attained at the match point.
show 6 more
The paper says, at Sec. IV.B
Pith says: minor objection 3 of 8
The symbol k denotes both the MPC time step in Eq. (8) and the number of neighbors in the weight notation w_knn of Eq. (9); please rename one of the two to avoid confusion.
The paper says, at Fig. 6
Pith says: minor objection 4 of 8
The perturbation table is difficult to parse; please label the rows with the perturbation value, the columns with the controller variant, and use a consistent color scale so that the success rates are legible.
The paper says, at Sec. VII and Fig. 7
Pith says: minor objection 5 of 8
Please state how the representative hardware run was chosen and whether the four additional reruns share the same initial condition, since this determines how much generality the hardware claim carries.
The paper says, at Sec. V.B and Sec. VIII
Pith says: minor objection 6 of 8
The abstract's '1 kHz' is the GPU-inference bound from Table I, whereas the deployed hardware controller runs at 250 Hz; please make that distinction explicit.
The paper says, at Sec. VII
Pith says: minor objection 7 of 8
Please state whether the deployed pipeline includes the ancillary controller and the safety filter of Fig. 1, and whether the safety filter is active during the hardware runs.
The paper says, at Sec. IV.C and Fig. 5a
Pith says: minor objection 8 of 8
Section IV.C states that the iterative procedure runs for 1441 iterations while Fig. 5a shows only 900; please state whether the figure truncates the run and whether the region-of-attraction evaluation in Fig. 3 uses the final iteration-1441 network.
What would settle it
Run DAgger and MC-dropout for the same total expert-query budget as the full SIFS run in Section IV.C (1,189,320 aggregated points, 1441 iterations), or until their failure rates plateau. Compare final failure rates, final success sets, and the query/time at which each method first reaches SIFS's final failure rate. If either baseline reaches a final failure rate at or below SIFS's, the abstract's 'matches standard dataset aggregation performance with less than half the training data' is false and should be rewritten as 'reaches a given intermediate failure rate faster.' If both plateau above SIFS at the same budget, the data-efficiency claim survives.
from the skeptic's reading
How the review went
4 readings by deepseek-v4-flash · 13 Nov 2026
- Reader holds up, with conditions
- Skeptic disagrees
- Editor wrote the verdict above
- Referee major revision: 4 major objections, 8 minor
Not run on this paper: the simulated authors' defence.
Pith did not check this review's objections against the paper.
The review in full
The review in brief, machine-written
The paper's core idea—self-informed sampling (SIFS) using a jointly learned time-to-go as an informative-state oracle—is genuinely original. The T2G-based weighting for conflicting expert labels is also new and lightweight. The integrated architecture is a sensible assembly of existing building blocks, and the hardware validation on a Furuta pendulum, while limited, is a real step beyond the sim-only work that dominates this area. The authors are honest about limitations: they acknowledge T2G noise, suboptimal expert labels, and the remaining performance gap. The writing is clear and the method is coherently explained.
The soft spot is load-bearing. The stress-test note is correct: the data-efficiency comparison in Fig. 5b uses a one-sided stopping rule. DAgger and MC-dropout are run only until they match SIFS's failure rate, not until they converge or exhaust the same query budget. The paper never reports the baselines' final failure rates or plateaus. This means the claim that SIFS 'matches standard dataset aggregation performance with less than half the data and a third of the time' is not established by the evidence shown. It could be that SIFS converges to a worse final policy faster. The authors need to show the full learning curves for all methods, not just the match point.
Other concerns are real but less severe. The hardware demonstration is five runs of a single swing-up task—encouraging but not statistically robust. The safety filter's control-invariant set is never specified for the Furuta plant, so the formal constraint-satisfaction guarantee is unverified. Several hyperparameters are unreported. The abstract claims 1 kHz, but hardware runs at 250 Hz. These are addressable in revision.
Who this is for: researchers working on neural approximation of MPC, especially for time-optimal or computationally expensive formulations. The paper deserves a serious referee—the core idea is worth engaging with—but the referee should push hard on the stopping-rule issue and ask for full learning curves. I would not cite it in its current form, but I would read a revised version carefully.
The strongest objection
The central data-efficiency claim (abstract; b4; b169; conclusion) is that SIFS matches DAgger/MC-dropout with less than half the data and about one-third the training time. The only direct comparison is Fig. 5. Its panel (b) caption states that the reported queries and time are those 'needed to match the SIFS failure rate.' That is a one-sided stopping rule: the baselines are run only until they reach the failure level SIFS already has, not until they converge or until the same total query budget as the full SIFS run described in Section IV.C (1441 iterations, 1,189,320 aggregated points). If DAgger or MC-dropout keep improving beyond that match point, SIFS does not 'match' their performance; it merely reaches a particular intermediate failure rate faster. The paper does not report the baselines' final failure rates or identify the SIFS checkpoint used for the match, so the comparison cannot distinguish 'SIFS is more data-efficient' from 'SIFS converges to a worse final policy more quickly.' This is load-bearing because the novelty of SIFS rests on this comparison and because the baselines are the standard DAgger/MC-dropout methods, not weak controls. The text at b167-b169 reports a faster reduction and a match point, but never a baseline plateau. The reader's T2G-faithfulness concern is relevant to the mechanism, but even a faithful T2G would not rescue the claim if the baselines are truncated; the stopping rule must be fixed first.
The paper is a coherent engineering integration with a real hardware swing-up (b182, b195) and a clear simulation comparison (b149). The reader's CONDITIONAL verdict is reasonable. My single most load-bearing concern is not the learned T2G per se, but the empirical comparison that carries the headline SIFS claim. The paper's own Fig. 5b caption says the baselines are measured by the queries/time 'needed to match the SIFS failure rate.' That makes the comparison asymmetric: SIFS is the target, and DAgger/MC-dropout are stopped at the target. The abstract and conclusion, however, state that SIFS 'matches standard dataset aggregation performance' (b4; conclusion). Matching a target set by the proposed method is not the same as matching the baseline's best performance. If DAgger's failure rate continues to fall after the match point, the correct statement is that SIFS is faster to a given intermediate failure rate, not that it matches DAgger with less data. The full SIFS training run in Section IV.C is 1441 iterations and 1,189,320 aggregated points; Fig. 5b's x-axis goes only to 240 iterations, and the paper does not report the baselines at the full budget. This is a concrete, checkable gap. The reader's T2G-oracle concern (b100) is a legitimate mechanism-level worry, and the unspecified control-invariant set C in Section III.F is a real missing premise for the formal constraint-satisfaction sentence. Both are worth fixing. But neither is as directly dispositive for the headline data-efficiency number as the truncated baseline comparison: a faithful T2G would still leave the abstract's 'matches' unsupported if the baselines were not run to convergence, while a proper baseline run would settle the data-efficiency claim regardless of the T2G explanation. The 1 kHz abstract versus 250 Hz hardware discrepancy (b5 versus b195) is a scope and communication issue, not the load-bearing one. The appropriate verdict remains CONDITIONAL, with the condition that the baseline comparison be run to the same budget or to convergence; if the baselines then match or beat SIFS, the central claim needs to be rewritten. I therefore disagree with the reader's specific assessment that the data-efficiency claim is already plausibly supported, and I would keep the verdict at CONDITIONAL but for this different reason.
The referee's report
The target problem is genuinely harder than the quadratic-tracking MPC imitation of most prior work, and a complete architecture with real-hardware validation would be a useful contribution if the claims hold. I credit the paper's strengths explicitly: a hardware demonstration on a Furuta pendulum with a long-horizon, nonconvex, hard-terminal-constraint TOMPC expert is rare; Table I gives a concrete and favorable computation-time comparison; the component ablations in Figs. 4 and 6 are good practice; the authors acknowledge expert suboptimality, learned-T2G noise, and remaining performance gaps; and the project webpage with videos and implementation details aids reproducibility. The obstacles are the strength of the evidence behind the two headline claims (data efficiency and region-of-attraction parity), the unspecified safe set behind the formal constraint-satisfaction guarantee, and the heuristic justification of the T2G selection signal. These are fixable within the manuscript's scope, so the appropriate outcome is a major revision rather than a rejection.
Pith's reading of the paper (a machine's summary, not the authors')
This paper presents an integrated architecture for training a neural network to imitate a computationally expensive time-optimal model predictive controller (TOMPC) and deploying it in real time on a physical Furuta pendulum. The key claim is that the proposed self-informed sampling (SIFS) strategy, which uses the network's own learned time-to-go prediction to identify states where the expert controller should be queried, matches the performance of standard data aggregation (DAgger) with less than half the training data and about one-third the training time. The resulting neural controller achieves a 91.5% success rate over 1000 random initial states, matching the expert's region of attraction, while raising the control frequency from 1.4 Hz to up to 1 kHz. The paper further demonstrates real-world swing-up and stabilization on hardware, with trajectories closely matching simulation. If correct, this work shows that neural imitation can make long-horizon, nonconvex TOMPC with hard terminal constraints deployable on fast dynamical systems, overcoming the data-generation bottleneck that previously limited such approaches.
What carries the argument. The central mechanism is the jointly learned time-to-go (T2G) output head of the neural network. The T2G is defined as the first time the predicted MPC trajectory from a state lies within a tolerance of the target state. Along a time-optimal trajectory, the T2G decreases by exactly one sampling interval per step. The self-informed sampling (SIFS) algorithm monitors the T2G trend over entire rollouts: if the T2G deviates persistently from this expected linear decrease (e.g., stagnation, increases, or irregular jumps), the states at that transition are flagged as informative and the expert MPC is queried there. This serves as a direct, dynamics-aware indicator of control quality, unlike indirect proxies such as epistemic uncertainty or trajectory distance. The paper also employs a lightweight T2G-based weighting scheme for conflicting expert labels (Section III-C), optimal sampling to reduce dataset correlation (Section III-B), a safety filter with logarithmic barrier and quadratic penalty in the loss (Section III-F), and ancillary controllers (DIC and LQR) for plant-model mismatch compensation (Section III-E).
Core claim. The paper claims that a neural network can be trained to imitate a long-horizon, nonconvex time-optimal model predictive controller (TOMPC) with hard terminal constraints, and that the resulting neural network model predictive controller (NNMPC) can be deployed in real time on a fast dynamical system (a Furuta pendulum) while matching the expert controller's region of attraction. The central discovery is that self-informed sampling (SIFS), which uses a jointly learned time-to-go (T2G) value function to identify states where the policy deviates from time-optimal behavior, drastically reduces the amount of expert data required for training. The authors argue that a T2G prediction that decreases by exactly one sampling interval per time step is unlikely to occur by chance unless the network has learned both the T2G of visited states and the controls that steer the system to the correct next state, making it a direct and reliable indicator of control quality. The paper reports that SIFS matches standard dataset aggregation (DAgger) performance in roughly one-third the training time and with less than half the training data, and that the resulting NNMPC achieves a 91.5% success rate over 1000 sampled initial states (identical to the expert MPC), while raising the control frequency from 1.4 Hz to up to 1 kHz. On hardware, the controller swings up and stabilizes the pendulum along trajectories closely matching simulation.
The premise it rests on. The self-informed sampling (SIFS) strategy assumes that a neural network that cannot decrement an internal counter will only produce a time-to-go that decreases by exactly one sampling interval per step if it has genuinely learned the T2G of visited states and predicts controls that steer to the correct next state; if the learned T2G is a noisy or biased oracle, the criterion will flag the wrong states and the data-efficiency claim collapses.
If the paper is right (Pith's inferences, not the authors')
Neural imitation of time-optimal MPC with hard terminal constraints becomes feasible for fast dynamical systems where the expert's computational cost previously limited both online deployment and training-data generation.
The SIFS criterion can be applied to any control problem with a natural value function that decreases linearly along optimal trajectories, potentially extending beyond time-optimal objectives to other monotonically decreasing cost-to-go functions.
The integrated architecture demonstrates that dataset correlation, suboptimal labels, covariate shift, constraint satisfaction, and model mismatch can be jointly addressed in a single training and deployment pipeline, reducing the fragmentation of existing NNMPC approaches.
The 700-fold reduction in worst-case computation time (from 732 ms to 1 ms) opens the door to deploying neural approximations of complex MPC formulations on embedded hardware with limited computational resources.
The SIFS criterion's reliance on a learned T2G that decreases by exactly one sampling interval per step could be adapted to other iterative optimal control problems where the cost-to-go satisfies a Bellman-like recursion, not just time-optimal formulations.
The paper's hardware validation on a Furuta pendulum, while convincing, is a single test case; the scalability of SIFS to higher-dimensional systems (e.g., quadrotors, manipulators) where the T2G may be harder to learn or where the state space is larger remains to be demonstrated.
The authors acknowledge that conflicting labels from equally time-optimal modes remain unresolved by their weighting scheme; a future extension could use mixture density networks or multi-modal architectures to explicitly handle such cases.
The safety filter relies on a control-invariant safe set C that is not explicitly specified for the Furuta plant; the practical constraint-satisfaction guarantee therefore depends on the conservativeness of the chosen CBF or invariant set, which may limit performance in more complex environments.
Check this review
Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.
How often reviews like this one are right
Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.
How often objections from reviews like this one are right: not yet measured
Reviews by people
This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.
The paper
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) 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.
The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.
abstract compiled · 2026-11-13
The authors' main claim, as the classifier picked it out 0.84
“The resulting controller matches the region of attraction of the expert TOMPC while raising the control frequency from 1.4~Hz to 1~kHz.”
How the main claim is stated: asserted 1.00
What the abstract itself states, as probabilities: claims first 0.08 · states limitation 0.41 · human participants 0.03 · controlled comparison 0.81 · randomized assignment 0.01 · states quantitative result 0.99 · compares with prior results 0.92 · mentions code or data availability 0.03
The abstract, sentence by sentence
This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.
| Sentence | Role | The authors' own work | States the main conclusion | Strength |
|---|---|---|---|---|
| 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. | background 0.99 | 0.05 | 0.03 | asserted 1.00 |
| 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. | gap 0.86 | 0.05 | 0.03 | asserted 0.97 |
| This work imitates a long-horizon nonconvex TOMPC with hard terminal constraints, whose high computational cost limits training-data generation. | background 0.38 | 0.63 | 0.03 | asserted 1.00 |
| 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. | contribution 0.99 | 0.96 | 0.30 | asserted 1.00 |
| SIFS matches standard dataset aggregation performance in roughly one-third of the training time and with less than half the training data. | result 1.00 | 0.95 | 0.13 | asserted 1.00 |
| The resulting controller matches the region of attraction of the expert TOMPC while raising the control frequency from 1.4~Hz to 1~kHz. | result 1.00 | 0.95 | 0.34 | asserted 1.00 |
| On hardware, the controller swings up and stabilizes the pendulum along trajectories closely matching the expert TOMPC solution. | result 1.00 | 0.95 | 0.23 | asserted 1.00 |
What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.99
What was examined, and what was not
Sources read
- arXiv listing abstract v1
Not examined
- full text
- figures and tables
- references and citing sentences
- relations to other papers
Levels run
- P abstract
Not run
- S sections
- T body sentences
- C citations
- R cross-paper relations
How the labels were made
Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-11-13: 1 request, 5,739 input tokens.
Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON
Figures (7)
Figures are the authors', reproduced from the paper.
References (35)
- [1] R. Verschueren, H. J. Ferreau, A. Zanariniet al., “A stabilizing nonlinear model predictive control scheme for time-optimal point-to- point motions,” in2017 IEEE 56th Annual Conference on Decision and Control (CDC), 2017, pp. 2525–2530 on Pith
- [2] A. Florez, W. Decr ´e, and J. Swevers, “Practical approaches for time- optimal mpc: Dual stage and artificial reference methods,” in2025 European Control Conference (ECC), 2025, pp. 358–363 on Pith
- [3] D. Dirckx, M. Bos, B. Vandewalet al., “Asap-mpc: an asynchronous update scheme for online motion planning with nonlinear model predictive control,”Auton. Robots, vol. 49, no. 1, Mar. 2025 on Pith
- [4] M. Bain and C. Sammut, “A framework for behavioural cloning,” in Machine Intelligence 15, Intelligent Agents [St. Catherine’s College, Oxford, July 1995]. GBR: Oxford University, 1999, p. 103–129 on Pith
- [5] B. Karg and S. Lucia, “Efficient representation and approximation of model predictive control laws via deep learning,”IEEE Transactions on Cybernetics, vol. 50, no. 9, pp. 3866–3878, 2020 on Pith
- [6] N. R. Mohanty, P. Ubare, V . Patneet al., “Embedded dnn-based non- linear mpc with guaranteed constraint satisfaction,”Neural Computing and Applications, vol. 37, no. 18, p. 12305–12328, Mar. 2025 on Pith
- [7] T. Li and K. You, “Neural mpc for quadrotor trajectory tracking on embedded gpus,”Control Theory and Technology, vol. 24, pp. 271– 281, 2026 on Pith
- [8] A. Tagliabue and J. P. How, “Efficient deep learning of robust policies from mpc using imitation and tube-guided data augmentation,”IEEE Transactions on Robotics, vol. 40, pp. 4301–4321, 2024 on Pith
- [9] C. Kong, B. Wang, L. Heet al., “Ann-mpc: An adaptive neural network-based nonlinear model predictive control method for robust trajectory tracking of quadrotors,”IEEE Robotics and Automation Letters, vol. 10, no. 10, pp. 10 966–10 973, 2025 on Pith
- [10] H. Hose, A. Gr ¨afe, and S. Trimpe, “Parameter-adaptive approxi- mate MPC: Tuning neural-network controllers without retraining,” in Proceedings of the 6th Annual Learning for Dynamics & Control Conference, ser. Proceedings of Machine Learning Research, A. Abate, M. Cannon, K.... on Pith
- [11] S. Ross, G. Gordon, and D. Bagnell, “A reduction of imitation learning and structured prediction to no-regret online learning,” in Proceedings of the Fourteenth International Conference on Artificial Intelligence and Statistics, ser. Proceedings of Machine Learning Research, G... on Pith
- [12] A. Castillo, E. Pryor, A. El Fathiet al., “Neural networks for on- chip model predictive control: A method to build optimized training datasets and its application to type-1 diabetes,”IEEE Transactions on Cybernetics, vol. 55, no. 7, pp. 3062–3072, 2025 on Pith
- [13] S. Zhang, Z. Cao, D. Sadighet al., “Confidence-aware imitation learning from demonstrations with varying optimality,” inAdvances in Neural Information Processing Systems, M. Ranzato, A. Beygelzimer, Y . Dauphinet al., Eds., vol. 34. Curran Associates, Inc., 2021, pp. 12 340–12 350 on Pith
- [14] A. Bemporad, M. Morari, V . Duaet al., “The explicit linear quadratic regulator for constrained systems,”Automatica, vol. 38, no. 1, pp. 3– 20, 2002 on Pith
- [15] Y . Zhang, Y . Xie, H. Liuet al., “Scizor: A self-supervised approach to data curation for large-scale imitation learning,” arXiv:2505.22626, 2025, preprint on PithPith's reviewarXiv
- [16] M. Yang, S. Levine, and O. Nachum, “TRAIL: Near-optimal imita- tion learning with suboptimal data,” inInternational Conference on Learning Representations, 2022 on Pith
- [17] H. Xu, X. Zhan, H. Yinet al., “Discriminator-weighted offline imita- tion learning from suboptimal demonstrations,” inProceedings of the 39th International Conference on Machine Learning, ser. Proceedings of Machine Learning Research, K. Chaudhuri, S. Jegelka, L. Song et al., ... on Pith
- [18] M. Laskey, J. Lee, R. Foxet al., “Dart: Noise injection for robust imitation learning,” inProceedings of the 1st Annual Conference on Robot Learning, ser. Proceedings of Machine Learning Research, S. Levine, V . Vanhoucke, and K. Goldberg, Eds., vol. 78. PMLR, 13–15 Nov 2017, ... on Pith
- [19] Y . Cui, D. Isele, S. Niekumet al., “Uncertainty-aware data aggregation for deep imitation learning,” in2019 International Conference on Robotics and Automation (ICRA). IEEE Press, 2019, p. 761–767 on Pith
- [20] B. Lakshminarayanan, A. Pritzel, and C. Blundell, “Simple and scalable predictive uncertainty estimation using deep ensembles,” in Proceedings of the 31st International Conference on Neural Informa- tion Processing Systems, ser. NIPS’17. Red Hook, NY , USA: Curran Associates I... on Pith
show 15 more
- [21] Y . Gal and Z. Ghahramani, “Dropout as a bayesian approximation: Representing model uncertainty in deep learning,” inProceedings of The 33rd International Conference on Machine Learning, ser. Proceedings of Machine Learning Research, M. F. Balcan and K. Q. Weinberger, Eds., vo... on Pith
- [22] A. Giusti, M. Monsieur, and B. Mrak, “STRIDE: Investigating efficient sampling strategies for neural approximations of optimal control,” in Proceedings of the 23rd International Conference on Informatics in Control, Automation and Robotics (ICINCO), vol. 1, 2026, pp. 166– 173 on Pith
- [23] T. Zieger, A. Savchenko, T. Oehlschl ¨agelet al., “Towards safe neural network supported model predictive control,”IFAC-PapersOnLine, vol. 53, no. 2, pp. 5246–5251, 2020, 21st IFAC World Congress on Pith
- [24] S. Adhau, V . V . Naik, and S. Skogestad, “Constrained neural networks for approximate nonlinear model predictive control,” in2021 60th IEEE Conference on Decision and Control (CDC), 2021, pp. 295– 300 on Pith
- [25] A. D. Ames, X. Xu, J. W. Grizzleet al., “Control barrier function based quadratic programs for safety critical systems,”IEEE Transactions on Automatic Control, vol. 62, no. 8, pp. 3861–3876, 2017 on Pith
- [26] P. Christiano, Z. Shah, I. Mordatchet al., “Transfer from simulation to real world through learning deep inverse dynamics model,”arXiv preprint arXiv:1610.03518, 2016 on PithPith's reviewarXiv
- [27] J.-J. E. Slotine and W. Li,Applied Nonlinear Control. Englewood Cliffs, NJ: Prentice-Hall, 1991 on Pith
- [28] S. Kim, J. Lee, K. Hanet al., “Bridging the sim-to-real gap: A control framework for imitation learning of model predictive control,” International Journal of Control, Automation, and Systems, vol. 24, no. 3, p. 425–443, Feb. 2026 on Pith
- [29] K.-C. Hsu, H. Hu, and J. F. Fisac, “The safety filter: A unified view of safety-critical control in autonomous systems,”Annual Review of Control, Robotics, and Autonomous Systems, vol. 7, no. 1, pp. 47–72, 2024 on Pith
- [30] B. Cazzolato and Z. Prime, “On the dynamics of the furuta pendulum,” Journal of Control Science and Engineering, vol. 2011, 04 2011 on Pith
- [31] L. Vanroye, A. Sathya, J. De Schutteret al., “Fatrop: A fast constrained optimal control problem solver for robot trajectory optimization and control,” in2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2023, pp. 10 036–10 043 on Pith
- [32] A. Florez, A. Astudillo, W. Decr ´eet al., “Impact: A toolchain for nonlinear model predictive control specification, prototyping, and deployment,”IFAC-PapersOnLine, vol. 56, no. 2, p. 3164–3169, 2023 on Pith
- [33] T. S. Lembono, A. Paolillo, E. Pignatet al., “Memory of motion for warm-starting trajectory optimization,”IEEE Robotics and Automation Letters, vol. 5, no. 2, pp. 2594–2601, 2020 on Pith
- [34] M. Bos, W. Decr ´e, and J. Swevers, “ImpaC++t: A C++ toolchain for seamless deployment of nonlinear model predictive control with Impact,” Presented at the 45th Benelux Meeting on Systems and Control, Lommel, Belgium, Mar. 2026, abstract in the Book of Abstracts on Pith
- [35] M. Okamoto, J. Ren, Q. Maoet al., “Approximation of nonlinear model predictive control using mixture density networks,” inProc. Found. Comput.-Aided Process Oper./Chem. Process Control (FO- CAPO/CPC), 2023 on Pith
Challenges and discussion
Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.
Come back by email
A short list of the week's newly reviewed papers. Nothing is sent until you confirm from your inbox, and every email has a link that stops it. No account.
No comments yet.
Sign in to comment