Pith. sign in

REVIEW 3 major objections 5 minor 31 references

Stochastic Multiple Shooting Trajectory Optimization via Sequential Local Policy Evaluation

T0 review · 3 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Stochastic multiple shooting can hit terminal sets that single-shooting MPPI and CEM miss, by linking short segments with local feedback policies.

desk verdict Genuinely new stochastic multiple shooting idea with solid terminal-set results, but the sample-efficiency claim is undercut by rollout accounting that omits Jacobian synthesis and constraint verification. read the letter →

arxiv 2608.03978 v1 pith:QQDTDVXJ submitted 2026-08-04 cs.RO

Ashwin Gupta , Joseph Moore This is my paper
classification cs.RO
keywords stochasticmultipleshootingsampling-basedtrajectoryoptimizationterminalsetconvergencecross-entropymethodmodelpredictivepathintegralcontroltime-varyingLQRblack-boxdynamicsmodel-basedreinforcementlearning
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Sampling-based trajectory optimizers such as model predictive path integral control (MPPI) and the cross-entropy method (CEM) shoot the entire control sequence in one forward pass, which makes a distant terminal constraint hard to satisfy: small early errors amplify, so precise final sets demand many iterations and many rollouts. The paper argues that stochastic multiple shooting fixes this: split the horizon into short segments, optimize each segment's controls and starting state, and connect segments not by an exact state match but by a chance constraint that the segment ends inside the invariant set of the next segment's local feedback policy. A time-varying LQR cost-to-go supplies both the connection cost and the sampling distribution for the next segment's initial state, and the required Jacobians can be fit from rollouts, so no analytical model is needed. If the claim holds, high-accuracy maneuvers — pendulum swingup, post-stall precision landing — become reachable for sampling-based planners at a fraction of the sample cost, including when the dynamics are a learned neural network.

What carries the argument

The load-bearing object is the local time-varying LQR (TVLQR) feedback policy built around each optimized segment, whose backward-integrated Riccati matrix S_k defines a quadratic cost-to-go. A level set of that cost-to-go, the ellipsoid x^T S x = ρ, is treated as an invariant set of the closed-loop policy and does double duty: it is the target the preceding segment must reach (turning the defect equality constraint into the chance constraint p(F ∈ R_{j+1}) > 1 − ε_R), and its confidence-ellipsoid form Σ^{−1} = (Pχ²_n / ρ) S seeds the Gaussian used to sample the next segment's initial state. When the dynamics are black-box, the policy's Jacobians are recovered by a least-squares fit of pertu

What would settle it

Simulate the closed-loop TVLQR policy under the true stochastic dynamics from many states sampled on the boundary of the level-set ellipsoid R_{j+1} for a converged cartpole or quadplane solution, and count the trajectories that leave the ellipsoid before the planned handoff time. If the escape fraction is comparable to or larger than ε_R, the invariance property that justifies the relaxed defect constraint fails for the nonlinear system, and the reported terminal-set convergence rests on the Monte Carlo check alone.

Watch

Extended reading notes

Core claim

The paper's central claim is that the defect equality constraint that has kept multiple shooting out of reach for gradient-free stochastic optimizers can be relaxed into an inequality chance constraint: the terminal state of segment j must land, with probability at least 1 − ε_R, inside the invariant set R_{j+1} of the local time-varying LQR (TVLQR) policy built for the next segment. Optimization proceeds backward from a terminal segment that is driven into the goal set; every preceding segment minimizes its own running cost plus the quadratic cost-to-go V(x) = (x − x_{s_{n+1}})^T S_{s_{n+1}} (x − x_{s_{n+1}}) of the downstream policy, and its initial state is itself optimized, sampled from

Load-bearing premise

The load-bearing assumption is that the cost-to-go ellipsoid is a genuine invariant set of the feedback policy under the real nonlinear stochastic dynamics — a property the paper only proves for the locally linearized system, leaving the chance constraint backed by Monte Carlo rollout verification rather than a guarantee.

Editorial extensions

If this is right

  • Terminal-set convergence: on all three benchmarks only the multiple-shooting method enters the specified terminal set; single-shooting CEM and MPPI plateau at higher-cost trajectories outside it (123.4/130.3 → 23.8 on cartpole; 53.3/115.2 → 20.3 on the quadplane).
  • Sample efficiency: measured in full-horizon rollouts as the hardware-independent cost, the method reaches lower mean cost in fewer samples, and it optimizes shorter, better-conditioned segments rather than one long sensitive sequence.
  • Black-box applicability: least-squares Jacobian synthesis from perturbed rollouts provides the A_k, B_k matrices the feedback policy needs, which is what makes the method work on a non-differentiable neural-network dynamics model with process noise.
  • Generality: the framework is not tied to its two default ingredients — the paper states any stochastic optimizer and any feedback policy that carries a cost-to-go could populate the same segment-connection scheme.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: the backward pass that hands each preceding segment a cost-to-go target is structurally a policy-space dynamic program, which suggests the method may admit a convergence analysis as segment lengths shrink, and may ultimately be interpretable as block-coordinate descent on the full-horizon stochastic optimal control problem.
  • Editorial inference: the warm-start requirement (a few single-shooting iterations) makes receding-horizon operation the natural deployment setting, since each new solve already has a near-feasible trajectory from the previous step; the paper does not test this.
  • Editorial inference: because the join between segments is a set-membership chance constraint rather than an exact match, the scheme may tolerate model error in learned dynamics — a testable prediction is that terminal-set success degrades only gradually as process noise or model mismatch grows, up to the point where the ellipsoid stops being invariant.
  • Editorial inference: nothing in the connection scheme requires the segment targets to be the task's final goal set, so multi-phase maneuvers (swingup, then balance, then throw) could be chained in the same backward pass by declaring intermediate invariant sets.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes a stochastic multiple shooting trajectory optimization method. The trajectory is split into segments; each segment is optimized with cross-entropy optimization, and adjacent segments are connected by local time-varying LQR (TVLQR) feedback policies. Defect equality constraints are relaxed into chance constraints requiring the terminal state of a segment to lie in an invariant set of the next segment's policy. The paper also proposes a least-squares method to synthesize Jacobians from rollouts, enabling application to black-box dynamics. Experiments on cartpole, a learned cartpole model, and a VTOL quadplane are reported, with the multiple shooting method claimed to achieve better sample efficiency and the only method to converge to specified terminal sets.

Significance. If the claims hold, the paper addresses a genuine limitation of stochastic single shooting methods: poor terminal-constraint satisfaction over long horizons. The rollout-based Jacobian synthesis is practically useful for model-based RL with non-differentiable simulators. The empirical evaluations on underactuated, stochastic, and high-dimensional systems are relevant to robotics. The central idea of replacing defect equalities with invariant-set chance constraints is appealing. However, the sample-efficiency comparison is currently not fully supported because the reported rollout counts omit rollouts used for Jacobian synthesis and constraint verification, and the invariant-set assumption underlying Eq. (4) is not justified for nonlinear stochastic dynamics.

major comments (3)
  1. [Sec. V.A, Sec. IV-C, Sec. IV-F] The sample-efficiency claim is based on cost-vs-rollout curves in Figs. 4, 5, and 7, but the rollout counts appear to omit two classes of dynamics evaluations. First, the least-squares Jacobian synthesis in Sec. IV-C/Eq. (13) requires evaluating the dynamics at K perturbed state/input samples; for black-box models these are forward simulations. Second, Sec. IV-F verifies constraint satisfaction by rolling out a batch of full-horizon trajectories every O CEM iterations. The example in Sec. V.A counting only 3000 rollouts for multiple shooting does not include these overhead rollouts. Since 'improved sample efficiency' is a central contribution, the authors must either include all dynamics evaluations in the reported curves or explicitly justify why the omitted evaluations are negligible. The manuscript should also report K, O, the verification batch size, and the exact formulas used to co
  2. [Sec. IV-B, IV-D, IV-E, Eq. (4)] The defect relaxation in Eq. (4) treats the TVLQR cost-to-go level set R_{j+1} as an invariant set for the subsequent segment. This invariance is guaranteed, if at all, only for the local linearized system used to compute the LQR gains; for the original nonlinear stochastic dynamics the TVLQR cost-to-go ellipsoid is not generally invariant. The paper does not provide a local invariance theorem or an explicit statement that this is a heuristic assumption. Because the entire defect-relaxation argument rests on this invariance, the manuscript should either prove or clearly qualify this assumption. The Monte Carlo verification in Sec. IV-F can serve as empirical evidence for the terminal-set claim, but the formal language in Eq. (4) and Algorithm 1 should be brought in line with what is actually guaranteed.
  3. [Sec. IV-F, Table I] The probabilistic constraint satisfaction criterion is not specified quantitatively. Eq. (1) requires p(x_T in X_f) > 1 - epsilon, but the paper does not state the value of epsilon, the percentage of verification rollouts considered 'sufficient,' the batch size used in verification, or the number of trials used for the checkmarks in Table I. Without these parameters, the terminal-set convergence results are not quantitatively reproducible, and the comparison between methods in Table I cannot be fully assessed. This also compounds the rollout-accounting problem in the first major comment.
minor comments (5)
  1. [Sec. IV-B, Eq. (11)] In the text following Eq. (11), B_k is defined as the derivative with respect to x_k; this should be the derivative with respect to u_k. As written it duplicates the definition of A_k.
  2. [Alg. 1, line 10] The notation 'x_sf' appears to be a typo for 'x_sn' (the segment start state). The pseudocode would benefit from consistent indexing.
  3. [Sec. V] Hyperparameters such as CEM batch size, number of CEM iterations, covariance magnitudes, and the checkpoint frequency O are not reported for each experiment. The example in Sec. V.A uses illustrative numbers, but the actual values used to generate Figs. 4, 5, and 7 are missing. This limits reproducibility.
  4. [Sec. V.C] The process noise magnitudes for the VTOL quadplane are not specified. Since the method is explicitly stochastic, reporting these values is important for reproducibility and for interpreting the chance-constraint results.
  5. [Sec. V] Table I reports mean terminal costs but not standard deviations or the distribution across the 10 trials. Given the stochastic nature of the problem, adding spread or confidence intervals would strengthen the comparison.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the central multiple-shooting construction and terminal-set benchmarks are independent; only minor non-load-bearing self-citations appear.

full rationale

The paper's core contribution is an algorithmic construction, not a theorem derivation, and its central claims are tested against external baselines (MPPI and single-shooting CEM) on three systems with externally specified terminal sets. The defect relaxation in Eq. 4 replaces the equality defect with a chance constraint using local TVLQR feedback policies; the TVLQR cost-to-go level set is used as the invariant set. This is a design choice, not a fitted parameter renamed as a prediction. The terminal-set convergence results in Table I are empirical benchmark outcomes, and the reported terminal costs are computed from rollouts of the compared algorithms, not from quantities used to fit the method. Self-citations appear: [1] provides the quadplane dynamics model (Sec. V-C) and [26] is mentioned as related work (Sec. II-C). Neither is load-bearing: no uniqueness theorem is imported, no central premise reduces to these references, and no ansatz is smuggled in via self-citation. The paper's own limitation is explicit: it says 'we hypothesize that if the defects are small, the costs will be similar in magnitude' (Sec. IV) and notes the TVLQR cost-to-go is a valid contraction metric only 'for the local linear system' (Sec. IV-D). Thus the invariance of the ellipsoid for the nonlinear stochastic dynamics is assumed and verified by Monte Carlo, which is a correctness risk, not circularity. The skeptic's rollout-count critique (Sec. V.A omitting Jacobian-synthesis and constraint-verification rollouts) is an evaluation accounting flaw affecting the sample-efficiency claim; it is not a circularity per the reviewing rules. Overall, no equation reduces to its own input, and no prediction is statistically forced by construction.

Assumptions & free parameters 8 free parameters · 4 assumptions · 0 invented entities

The central claim rests on an unproven invariance assumption (the TVLQR ellipsoid), a hardware-agnostic rollout metric that may omit Jacobian/verification samples, and a large set of unreported hyperparameters (rho, batch sizes, K, TVLQR weights, VTOL noise). These are the price the reader pays beyond the algorithm itself; no new physical entities are introduced.

free parameters (8)
  • Per-segment cost-to-go level-set thresholds rho_j
    Required inputs in Algorithm 1 (rho_{1:M}); define the size of the invariant-set ellipsoid for each segment's feedback policy. Values are never reported; they directly control whether a segment endpoint counts as satisfying the defect constraint and the covariance of initial-state samples (Sec. IV-D).
  • TVLQR weight matrices Q_c, R_c, Q_f
    Hand-chosen cost weights for the local feedback policy and cost-to-go; not reported in the paper; determine shape of the ellipsoids and gains.
  • CEM elite fraction = 0.05 for cartpole, unreported for VTOL
    Cross-entropy method's elite percentage; 5% stated for the cartpole experiments; not stated for VTOL (Sec. V-A, V-C).
  • CEM batch size and inner iteration count
    Number of samples per CEM iteration and maximum inner iterations are not reported; they affect sample efficiency claims (Sec. IV-F, Sec. V).
  • Initial state and control covariances Sigma_x, Sigma_u
    Exploration covariances for CEM; not reported; the paper only describes level-set-based Sigma_x for initial states (Sec. IV-D).
  • Number of least-squares samples K
    K perturbed states/controls used for Jacobian synthesis (Sec. IV-C); not reported; extra rollouts from these samples are not counted in the rollout-comparison metric.
  • Confidence percentile P for chi-squared ellipsoid
    The P-th percentile in Sec. IV-D is not specified; it indirectly sets the covariance scale for initial-state sampling.
  • Process noise standard deviations for VTOL
    Cartpole noise is given (0.1 m/s^2, 0.05 rad/s^2); VTOL noise is only described qualitatively as 'in acceleration and angular acceleration' (Sec. V-C).
assumptions (4)
  • ad hoc to paper The TVLQR cost-to-go level set is a valid invariant set for the nonlinear stochastic system under the local feedback policy.
    Introduced in Sec. IV and Sec. IV-D to justify replacing the defect equality constraint with Eq. 4. Not proven for the nonlinear stochastic dynamics; the paper only notes the cost-to-go is a contraction metric for the local linear system.
  • domain assumption Chance constraints (p(x_T in X_f) > 1-epsilon) can be evaluated reliably from a finite batch of Monte Carlo rollouts.
    Used in Sec. IV-F as the constraint satisfaction criterion during CEM optimization; no confidence interval on the probability estimate is provided.
  • ad hoc to paper The total number of full-horizon rollouts is a fair, hardware-agnostic proxy for compute time when comparing single shooting and multiple shooting.
    Defined in Sec. V; however, extra rollouts used for Jacobian synthesis (Sec. IV-C) and for constraint verification (Sec. IV-F) are not explicitly counted, so the proxy may not be fair.
  • standard math Cross-entropy method with multivariate Gaussian distributions converges to a near-optimal solution for each segment problem.
    Used as the underlying stochastic optimizer (Sec. IV-A); standard but with no convergence guarantee for the non-convex segment problems.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Stochastic Multiple Shooting Trajectory Optimization via Sequential Local Policy Evaluation." pith.science (2026). https://pith.science/paper/QQDTDVXJ

@misc{pith2026260803978,
  author       = {Pith},
  title        = {Pith review of: Stochastic Multiple Shooting Trajectory Optimization via Sequential Local Policy Evaluation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QQDTDVXJ}},
  note         = {Machine review of arXiv:2608.03978}
}
read the original abstract

Stochastic single shooting trajectory optimization methods such as Model Predictive Path Integral control (MPPI) have been widely adopted in robotics due to their ability to reason about probabilistic dynamics and provide solutions where model gradients are noisy, costly to evaluate, or unavailable. However, satisfaction of terminal constraints when shooting over long action sequences is often sample inefficient, requiring a large number of iterations for convergence. In this paper, we present a stochastic multiple shooting method that optimizes short control action sequences connected via local feedback policies to improve sample efficiency and convergence to a terminal set. Additionally, we show that we are able to synthesize approximate system Jacobians purely from rollouts, making the method suitable for model-based reinforcement learning with black-box dynamics. We demonstrate the algorithm has improved sample efficiency and terminal set convergence for three nonlinear, underactuated optimization problems: a classic cartpole swingup task with analytical dynamics, a cartpole swingup task with learned neural network dynamics, and a VTOL quadplane performing a high angle-of-attack, precision post-stall landing maneuver.

Figures

Figures reproduced from arXiv: 2608.03978 by the authors.

Figure 1
Figure 1. Illustration of the multiple shooting approach in state space on [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of the multiple shooting framework. Blue dots indicate the intermediate states [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 4
Figure 4. Number of rollouts vs mean cost for 10 trials of the cartpole swing-up problem using each optimization algorithm. Error bars indicate one standard deviation [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figures from the paper (5 more)
Figure 3
Figure 3. Figure 3: Cartpole Swing-Up Trajectories. The multiple shooting method [PITH_FULL_IMAGE:figures/full_fig_p006_3.png]
Figure 6
Figure 6. Figure 6: Visualization of quadplane precision landing trajectory. The system [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 5
Figure 5. Figure 5: Number of rollouts vs mean cost for 10 trials of the cartpole swing [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 7
Figure 7. Figure 7: Number of rollouts vs mean cost for 10 trials of the VTOL landing [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]
Figure 8
Figure 8. Figure 8: Quadplane Trajectories. Multiple shooting enables convergence into [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

31 extracted references · 24 canonical work pages

  1. [1]

    Agile Fixed-Wing UA Vs for Urban Swarm Operations,

    M. Basescu, A. Polevoy, B. Yeh, L. Scheuer, E. Sutton, and J. Moore, “Agile Fixed-Wing UA Vs for Urban Swarm Operations,”IEEE Trans- actions on Field Robotics, vol. 1, pp. 394–423, Jan. 2024

    work page 2024

  2. [2]

    Data Efficient Reinforcement Learning for Legged Robots

    Y . Yang, K. Caluwaerts, A. Iscen, T. Zhang, J. Tan, and V . Sindhwani, “Data efficient reinforcement learning for legged robots,” 2019. [Online]. Available: https://arxiv.org/abs/1907.03613

    work page Pith review arXiv 2019

  3. [3]

    A multiple shooting algorithm for direct solution of optimal control problems*,

    H. Bock and K. Plitt, “A multiple shooting algorithm for direct solution of optimal control problems*,”IFAC Proceedings Volumes, vol. 17, no. 2, pp. 1603–1608, 1984, 9th IFAC World Congress: A Bridge Between Control Science and Technology, Budapest, Hungary, 2-6 July 1984

    work page 1984

  4. [4]

    On the implementation of an interior-point filter line-search algorithm for large-scale nonlinear programming,

    A. W ¨achter and L. Biegler, “On the implementation of an interior-point filter line-search algorithm for large-scale nonlinear programming,” Mathematical programming, vol. 106, pp. 25–57, 03 2006

    work page 2006

  5. [5]

    Model predictive path integral control using covariance variable importance sampling,

    G. Williams, A. Aldrich, and E. A. Theodorou, “Model predictive path integral control using covariance variable importance sampling,”CoRR, vol. abs/1509.01149, 2015. [Online]. Available: http://arxiv.org/abs/1509.01149

    arXiv 2015

  6. [6]

    Cross-entropy motion planning,

    M. Kobilarov, “Cross-entropy motion planning,”International Journal of Robotics Research, vol. 31, no. 7, Jun. 2012. [Online]. Available: https://doi.org/10.1177/0278364912444543

    doi:10.1177/0278364912444543 2012

  7. [7]

    Survey of numerical methods for trajectory optimization,

    J. T. Betts, “Survey of numerical methods for trajectory optimization,” Journal of guidance, control, and dynamics, vol. 21, no. 2, pp. 193– 207, 1998

    work page 1998

  8. [8]

    The trajectory bundle method: Unifying sequential- convex programming and sampling-based trajectory optimization,

    K. Tracy, J. Z. Zhang, J. Arrizabalaga, S. Schaal, Y . Tassa, T. Erez, and Z. Manchester, “The trajectory bundle method: Unifying sequential- convex programming and sampling-based trajectory optimization,” in Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), 2026

    work page 2026

Show all 31 references
  1. [9]

    On converting optimal control problems into nonlinear pro- gramming problems,

    D. Kraft, “On converting optimal control problems into nonlinear pro- gramming problems,” inComputational mathematical programming. Springer, 1985, pp. 261–280

    1985

  2. [10]

    A multiple shooting algorithm for direct solution of optimal control problems,

    H. G. Bock and K.-J. Plitt, “A multiple shooting algorithm for direct solution of optimal control problems,”IFAC Proceedings Volumes, vol. 17, no. 2, pp. 1603–1608, 1984

    1984

  3. [11]

    Evaluating direct transcription and nonlinear optimization methods for robot motion planning,

    D. Pardo, L. M ¨oller, M. Neunert, A. W. Winkler, and J. Buchli, “Evaluating direct transcription and nonlinear optimization methods for robot motion planning,”IEEE Robotics and Automation Letters, vol. 1, no. 2, pp. 946–953, 2016

    2016

  4. [12]

    Direct collocation methods for trajectory optimization in constrained robotic systems,

    R. Bordalba, T. Schoels, L. Ros, J. M. Porta, and M. Diehl, “Direct collocation methods for trajectory optimization in constrained robotic systems,”IEEE Transactions on Robotics, vol. 39, no. 1, p. 183–202, Feb. 2023. [Online]. Available: http://dx.doi.org/10.1109/TRO.2022.3193776

    2023

  5. [13]

    Direct trajectory optimization using nonlinear programming and collocation,

    C. HARGRA VES and S. Paris, “Direct trajectory optimization using nonlinear programming and collocation,”AIAA J. Guidance, vol. 10, pp. 338–342, 07 1987

    1987

  6. [14]

    A survey of numerical methods for optimal control,

    A. Rao, “A survey of numerical methods for optimal control,”Ad- vances in the Astronautical Sciences, vol. 135, 01 2010

    2010

  7. [15]

    A second-order gradient method for determining optimal trajectories of non-linear discrete-time systems,

    D. Q. Mayne, “A second-order gradient method for determining optimal trajectories of non-linear discrete-time systems,”International Journal of Control, vol. 3, pp. 85–95, 1966

    1966

  8. [16]

    ALTRO: A fast solver for constrained trajectory optimization,

    T. Howell, B. Jackson, and Z. Manchester, “ALTRO: A fast solver for constrained trajectory optimization,” inProceedings of (IROS) IEEE/RSJ International Conference on Intelligent Robots and Systems, November 2019, pp. 7674 – 7679

    2019

  9. [17]

    A unified perspective on multiple shooting in differential dynamic programming,

    H. Li, W. Yu, T. Zhang, and P. M. Wensing, “A unified perspective on multiple shooting in differential dynamic programming,” 2023. [Online]. Available: https://arxiv.org/abs/2309.07872

    2023 arXiv

  10. [18]

    A multiple-shooting differential dynamic programming algorithm. part 1: Theory,

    E. Pellegrini and R. P. Russell, “A multiple-shooting differential dynamic programming algorithm. part 1: Theory,”Acta Astronautica, vol. 170, pp. 686–700, 2020. [Online]. Available: https://www.sciencedirect.com/science/article/pii/S0094576519314705

    2020

  11. [19]

    A. E. Bryson,Applied optimal control: optimization, estimation and control. Routledge, 2018

    2018

  12. [20]

    A generalized iterative LQG method for locally-optimal feedback control of constrained nonlinear stochastic systems,

    E. Todorov and W. Li, “A generalized iterative LQG method for locally-optimal feedback control of constrained nonlinear stochastic systems,” inProceedings of the 2005, American control conference, 2005.IEEE, 2005, pp. 300–306

    2005

  13. [21]

    Tube stochastic optimal control for nonlinear constrained trajectory optimization problems,

    N. Ozaki, S. Campagnola, and R. Funase, “Tube stochastic optimal control for nonlinear constrained trajectory optimization problems,” Journal of Guidance, Control, and Dynamics, vol. 43, no. 4, pp. 645– 655, 2020

    2020

  14. [22]

    Direct policy optimization using deterministic sampling and collocation,

    T. A. Howell, C. Fu, and Z. Manchester, “Direct policy optimization using deterministic sampling and collocation,”IEEE Robotics and Automation Letters, vol. 6, no. 3, pp. 5324–5331, 2021

    2021

  15. [23]

    Variational inference MPC using Tsallis divergence,

    Z. Wang, O. So, J. Gibson, B. Vlahov, M. S. Gandhi, G.-H. Liu, and E. A. Theodorou, “Variational inference MPC using Tsallis divergence,” 2021. [Online]. Available: https://arxiv.org/abs/2104.00241

    2021 arXiv

  16. [24]

    Robust model predictive path integral control: Analysis and performance guarantees,

    M. S. Gandhi, B. Vlahov, J. Gibson, G. Williams, and E. A. Theodorou, “Robust model predictive path integral control: Analysis and performance guarantees,”IEEE Robotics and Automation Letters, vol. 6, no. 2, pp. 1423–1430, 2021

    2021

  17. [25]

    Risk-aware model predictive path integral control using conditional value-at-risk,

    J. Yin, Z. Zhang, and P. Tsiotras, “Risk-aware model predictive path integral control using conditional value-at-risk,” 2022

    2022

  18. [26]

    Probably approximately cor- rect nonlinear model predictive control (PAC-NMPC),

    A. Polevoy, M. Kobilarov, and J. Moore, “Probably approximately cor- rect nonlinear model predictive control (PAC-NMPC),”IEEE Robotics and Automation Letters, vol. 8, no. 11, pp. 7226–7233, 2023

    2023

  19. [27]

    CoVO-MPC: Theoretical analysis of sampling-based MPC and optimal covariance design,

    Z. Yi, C. Pan, G. He, G. Qu, and G. Shi, “CoVO-MPC: Theoretical analysis of sampling-based MPC and optimal covariance design,” 2024

    2024

  20. [28]

    Model-based diffusion for trajectory optimization,

    C. Pan, Z. Yi, G. Shi, and G. Qu, “Model-based diffusion for trajectory optimization,” 2024. [Online]. Available: https://arxiv.org/abs/2407.01573

    2024 arXiv

  21. [29]

    CVXOPT: Python software for convex optimization,

    M. S. Andersen, J. Dahl, L. Vandenbergheet al., “CVXOPT: Python software for convex optimization,”URL https://cvxopt. org, vol. 64, 2013

    2013

  22. [30]

    Conic optimization via operator splitting and homogeneous self-dual embedding,

    B. O’Donoghue, E. Chu, N. Parikh, and S. Boyd, “Conic optimization via operator splitting and homogeneous self-dual embedding,”Journal of Optimization Theory and Applications, vol. 169, no. 3, pp. 1042–1068, June 2016. [Online]. Available: http://stanford.edu/ boyd/papers/scs.html

    2016

  23. [31]

    Control contraction metrics: Convex and intrinsic criteria for nonlinear feedback design,

    I. R. Manchester and J.-J. E. Slotine, “Control contraction metrics: Convex and intrinsic criteria for nonlinear feedback design,” 2017. [Online]. Available: https://arxiv.org/abs/1503.03144

    2017 arXiv

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.