Pith. sign in

REVIEW 3 major objections 4 minor 27 references

Safe and Performant Controller Synthesis using Gradient-based Model Predictive Control and Control Barrier Functions

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper claims that pairing gradient-based MPC with a CBF-QP safety filter yields controllers that are both provably safe and high-performance.

desk verdict A sensible engineering combination that overclaims formal safety; the empirical results are plausible but the safety certificate is not established. read the letter →

arxiv 2507.13872 v1 pith:FKF2NNE4 submitted 2025-07-18 eess.SY cs.ROcs.SY

Aditya Singh , Aastha Mishra , Manan Tayal , Shishir Kolathaya , Pushpak Jagtap This is my paper
classification eess.SYcs.ROcs.SY
keywords controlbarrierfunctionsmodelpredictivegradient-basedoptimizationL-BFGSsafetyfilteringquadrotornavigationunicyclecollisionavoidance
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

The paper proposes a two-stage control pipeline that separates performance from safety. First, a gradient-based MPC planner using L-BFGS optimizes a cost in which safety constraints appear as penalties rather than hard constraints, making the optimization fast and avoiding feasibility failures. Second, a Control Barrier Function-based Quadratic Program takes that planner's command as a reference and enforces hard safety with minimal deviation. The central claim is that because the reference controller is already safety-aware, the filter intervenes less, so the result is both formally safe and performs better than filtering an unsafe nominal controller. The paper supports this with simulations of a unicycle and a planar quadrotor, where the proposed method matches the 100% safety rate of the MPPI-CBF baseline while lowering cumulative cost and computation time.

What carries the argument

The load-bearing mechanism is the two-stage loop in Algorithm 1: at each time step, L-BFGS minimizes the soft-constrained MPC cost $C(x,u) = \sum_k r(x_k,u_k) + \phi(x_K) + \lambda \max\{0, -l(x_k)+\delta\}$, then the first control is passed as the reference to the CBF-QP, which solves $\min_{u \in U} \|u - u_{\mathrm{mpc}}\|^2$ subject to $L_f h + L_g h\,u + \kappa(h) \ge 0$. A control barrier function is a function whose non-negative superlevel set is the safe set and whose derivative condition lets a feasible controller keep the state in that set. For the unicycle and the quadrotor, the paper uses higher-order CBFs so that constraints on position become constraints on the control inputs, giving the filter the authority to correct unsafe commands.

What would settle it

Simulate the quadrotor from initial states near multiple walls and record whether the four prioritized CBF-QPs remain feasible at every time step; any infeasible step or any trajectory that crosses a boundary refutes the claimed provable safety. For the unicycle, sample states near the obstacle and check whether $\sup_{u \in U}[L_f h + L_g h\,u + \kappa(h)] \ge 0$ holds on a dense grid of states.

Watch

Extended reading notes

Core claim

The paper's central claim is that the conservatism of CBF-QP safety filters comes less from the filter itself and more from the reference controller being oblivious to safety. When the reference is produced by a gradient-based MPC that already treats the safety constraint as a large penalty, the QP only needs small corrections, so the filtered controller stays close to the task-optimal trajectory. In the paper's two case studies this arrangement achieves a 100% safety rate, with the MPPI-CBF baseline's cost 47.6% higher on the unicycle and 110.6% higher on the quadrotor, while the gradient-based planner's computation time scales roughly linearly from the 3D unicycle to the 6D quadrotor.

Load-bearing premise

The safety guarantee rests on the assumptions that the hand-built higher-order barrier functions in Eq. (12) and Eq. (14) are valid control barrier functions and that the safety QP always has a feasible solution; the quadrotor's four-QP cascade lacks a formal certificate, and Eq. (14) uses the undefined symbol $\omega$, so this is asserted rather than proven.

Editorial extensions

If this is right

  • In both simulated navigation case studies, the GMPC-CBF controller achieves a 100% safety rate, matching the CBF-filtered baseline while producing a lower cumulative cost.
  • Gradient-based planning scales better than sampling: computation time roughly doubles from the 3D unicycle to the 6D quadrotor, whereas the MPPI baseline grows by a factor of 3 to 4.
  • By moving safety constraints into the cost as penalties, the first stage removes hard nonlinear state constraints and the feasibility failures those constraints cause in nonlinear MPC.
  • Because the CBF-QP theorem guarantees forward invariance of the safe set when the barrier condition and QP feasibility hold, the safety guarantee is carried by the second stage rather than by the soft penalties in the first stage.

Reading between the lines

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

  • The paper leaves implicit that the same two-stage separation should transfer to safety filters other than CBF-QP, such as Hamilton-Jacobi reachability filters; the crucial ingredient is a safety-aware reference that keeps the filter's intervention small.
  • The penalty weight $\lambda$ is likely a tuning dial tracing a safety-performance trade-off; the reported 100% safety results are demonstrations for chosen values, not a guarantee of safety for every $\lambda$.
  • A testable extension is to verify numerically that the hand-crafted higher-order CBFs in Eq. (12) and Eq. (14) satisfy the barrier condition on a dense grid of states; if the quadrotor's prioritized QP cascade can be infeasible, the provable-safety claim may need to be restricted to the unicycle.
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 / 4 minor

Summary. The paper proposes a two-stage controller synthesis framework. In the first stage, an MPC problem with safety constraints relaxed as soft penalties is solved by gradient-based L-BFGS optimization. In the second stage, the resulting nominal control is passed through a CBF-QP safety filter. The authors claim that this yields controllers that are both performant and provably safe, and they evaluate the method on unicycle and planar quadrotor navigation tasks, comparing against GMPC, MPPI, and MPPI-CBF baselines. The empirical tables report lower cumulative cost and 100% safety rate for the proposed GMPC-CBF method, with lower computation time than MPPI-based approaches.

Significance. If the safety claim were properly established, the two-stage architecture would be a useful practical contribution: soft-constrained gradient MPC followed by a CBF-QP is a natural way to reduce the conservatism of pure CBF filtering while avoiding the feasibility burden of hard-constrained nonlinear MPC. The paper includes a reasonable ablation against GMPC, MPPI, and MPPI-CBF and reports safety rates, costs, and computation times, which is good experimental practice. The scalability argument based on the reported computation times is also potentially interesting. However, the central 'provably safe' assertion is not supported by the analysis as written. The validity of the proposed CBFs is asserted rather than verified, the quadrotor implementation uses a prioritized QP cascade for which no safety theorem is provided, and the cost metric is not defined on equal footing across methods. These issues directly affect the two main claims of the paper.

major comments (3)
  1. [Section III-B, Eq. (12), Section IV-A] The unicycle safety filter is built on the function h in Eq. (12), which is asserted to be a valid HOCBF, but the conditions of Theorem 1 are never checked. In particular, L_g h = 2v(-x sinθ + y cosθ), so the control term vanishes on the set x sinθ = y cosθ; the paper gives no argument that L_f h + κ(h) ≥ 0 on that set, and therefore no argument that the CBF-QP in Eq. (10) is feasible at every state. Moreover, the safe set {x : h(x) ≥ 0} is not equivalent to the original constraint l(x) > 0; for a state just inside the obstacle radius with an outward heading, h can be positive, so forward invariance of {h ≥ 0} would not by itself imply collision avoidance. The 'provably safe' claim in the abstract thus rests on an unverified identification between the designed HOCBF and the actual safety specification.
  2. [Section IV-B, Eq. (14)] The quadrotor implementation replaces the single CBF-QP in Eq. (10) with a prioritized cascade of four QPs whose solutions are passed sequentially as reference inputs. This cascade is not covered by Theorem 1: no proof is given that applying the last QP's solution preserves the constraints of the earlier QPs, nor that the cascade is feasible whenever a single CBF-QP would be. In addition, the conditions C1–C4 in Eq. (14) use an undefined symbol ω, and no derivation is provided to show that these expressions are valid HOCBF conditions for the dynamics in Eq. (13) or that they are affine in the actual control inputs (F, M). As a consequence, the formal safety guarantee announced in the paper does not apply to the controller that is actually evaluated in the quadrotor experiments.
  3. [Section IV-A, Table I] The 'Cumulative Cost' metric is defined as the cost accumulated 'over the safe trajectories,' while the same table reports different safety rates across methods (85–88% for GMPC and MPPI versus 100% for the CBF-filtered methods). This makes the performance comparison between filtered and unfiltered methods ambiguous: excluding unsafe rollouts from the cost aggregate can bias the comparison in either direction. The paper should either report cost over all trajectories or clearly state the common set of trials over which costs are compared. In addition, the experimental section lacks the parameter values (α, λ, δ), obstacle geometry, number of trials, and MPPI settings needed to reproduce Table I.
minor comments (4)
  1. [Eq. (14)] The symbol ω is used without definition; presumably it denotes the angular velocity ẍθ from Eq. (13), but this should be stated explicitly along with which quantities are state-dependent.
  2. [Theorem 1 and Eq. (10)] The class-K function κ is never specified. The common choice κ(h) = αh should be stated, since the feasibility and safety arguments depend on its concrete form.
  3. [Problem 4] The penalty term λ max{0, -l(x) + δ} is not clearly connected to the strict constraint l(x) > 0 in Problem 1; the role of δ and its relation to the safety margin should be clarified.
  4. [Figure 2] The caption refers to walls, floor, and ceiling boundaries in what is described as a planar x–z quadrotor setup; the geometry of the simulated room should be described more precisely.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the safety guarantee is imported from external CBF theory, and the unverified HOCBF conditions are a correctness gap, not a circular reduction.

full rationale

The claimed derivation chain is: gradient-based MPC solves a soft-constrained problem; a CBF-QP filters the resulting reference using a hand-defined barrier h; safety then follows from the standard CBF forward-invariance theorem. Each link is externally grounded rather than circular. The theorem in Section III-B (Eqs. 8-10) is cited to the established CBF literature and is not derived from this paper's fitted values; its assumptions, namely a valid CBF and extended class-K function, are stated but not verified for the specific functions in Eq. (12) and Eq. (14). That missing verification is a correctness risk, not a circular reduction: the safety conclusion is not assumed as an input, and the barrier candidates are not defined in terms of the measured safety rates or cumulative costs. There are no fitted parameters renamed as predictions; cumulative cost, safety rate, and computation time are measured in simulation against external baselines. The self-citations in the reference list and conclusion appear in background or future-work statements and do not carry the safety argument; the load-bearing citations are to the external CBF and HOCBF literature. Therefore no step reduces by construction to its own inputs, and the paper is self-contained with respect to the claims it actually verifies empirically.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The safety claim is imported from the standard CBF-QP theorem rather than derived for these systems, and the validity and feasibility hypotheses are never checked. The three tuning parameters (λ, δ, α) are not reported, and the empirical evaluation lacks code and data, so the reader cannot verify the central result independently.

free parameters (3)
  • λ (safety penalty weight) = not reported
    Trade-off between safety penalty and task cost in Problem 4; no value or sensitivity analysis is given.
  • δ (safety margin) = not reported
    Inflates the soft safety constraint to enforce a strict inequality; value is not reported.
  • α (HOCBF convergence rate) = not reported
    Controls the rate of convergence to the safe set in Eq. (12); value is not reported.
assumptions (4)
  • standard math The standard CBF theorem (Theorem 1) from Ames et al. applies to the control-affine system and guarantees forward invariance if h is a valid CBF and the controller is Lipschitz.
    The paper's safety guarantee for Eq. (10) is inherited entirely from this cited theorem, presented in Section III-B.
  • ad hoc to paper The HOCBFs in Eq. (12) and Eq. (14) are valid CBFs for the unicycle and quadrotor systems respectively.
    No derivation or verification is provided; Eq. (14) contains an undefined ω and Eq. (12) omits the obstacle center in its formula.
  • domain assumption The CBF-QP in Eq. (10) is feasible at every state visited during simulation, including under the prioritized four-QP cascade used in Section IV-B.
    Feasibility is required for the safety guarantee but is neither proven nor discussed.
  • domain assumption L-BFGS converges to a sufficiently good local optimum of the nonconvex Problem 4 at every MPC step.
    The performance claims depend on the optimizer finding useful control sequences; no optimality guarantee exists for the nonconvex cost.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Safe and Performant Controller Synthesis using Gradient-based Model Predictive Control and Control Barrier Functions." pith.science (2026). https://pith.science/paper/FKF2NNE4

@misc{pith2026250713872,
  author       = {Pith},
  title        = {Pith review of: Safe and Performant Controller Synthesis using Gradient-based Model Predictive Control and Control Barrier Functions},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FKF2NNE4}},
  note         = {Machine review of arXiv:2507.13872}
}
read the original abstract

Ensuring both performance and safety is critical for autonomous systems operating in real-world environments. While safety filters such as Control Barrier Functions (CBFs) enforce constraints by modifying nominal controllers in real time, they can become overly conservative when the nominal policy lacks safety awareness. Conversely, solving State-Constrained Optimal Control Problems (SC-OCPs) via dynamic programming offers formal guarantees but is intractable in high-dimensional systems. In this work, we propose a novel two-stage framework that combines gradient-based Model Predictive Control (MPC) with CBF-based safety filtering for co-optimizing safety and performance. In the first stage, we relax safety constraints as penalties in the cost function, enabling fast optimization via gradient-based methods. This step improves scalability and avoids feasibility issues associated with hard constraints. In the second stage, we modify the resulting controller using a CBF-based Quadratic Program (CBF-QP), which enforces hard safety constraints with minimal deviation from the reference. Our approach yields controllers that are both performant and provably safe. We validate the proposed framework on two case studies, showcasing its ability to synthesize scalable, safe, and high-performance controllers for complex, high-dimensional autonomous systems.

Figures

Figures reproduced from arXiv: 2507.13872 by the authors.

Figure 1
Figure 1. Trajectories from a common initial state is shown, [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Trajectories originating from a common initial state are illustrated, where the dark grey region denotes the walls, floor, [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

27 extracted references · 13 canonical work pages

  1. [1]

    The safety filter: A unified view of safety-critical control in autonomous systems,

    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. V olume 7, 2024, pp. 47–72, 2024. [Online]. Available: https://www.annualreviews.org/content/journals/10.1146/annurev- control-071723-102940

    doi:10.1146/annurev- 2024

  2. [2]

    Data-driven safety filters: Hamilton-jacobi reachability, control barrier functions, and predictive methods for uncertain systems,

    K. P. Wabersich, A. J. Taylor, J. J. Choi, K. Sreenath, C. J. Tom- lin, A. D. Ames, and M. N. Zeilinger, “Data-driven safety filters: Hamilton-jacobi reachability, control barrier functions, and predictive methods for uncertain systems,” IEEE Control Systems Magazine , vol. 43, no. 5, pp. 137–177, 2023

    2023

  3. [3]

    Control barrier function based quadratic programs for safety critical systems,

    A. D. Ames, X. Xu, J. W. Grizzle, and P. Tabuada, “Control barrier function based quadratic programs for safety critical systems,” IEEE Transactions on Automatic Control , vol. 62, no. 8, pp. 3861–3876, 2017

    2017

  4. [4]

    Hamilton-Jacobi reachability: A brief overview and recent advances,

    S. Bansal, M. Chen, S. Herbert, and C. J. Tomlin, “Hamilton-Jacobi reachability: A brief overview and recent advances,” in 2017 IEEE 56th Annual Conference on Decision and Control (CDC) , 2017

    2017

  5. [5]

    On safety and liveness filtering using hamilton–jacobi reachability analysis,

    J. Borquez, K. Chakraborty, H. Wang, and S. Bansal, “On safety and liveness filtering using hamilton–jacobi reachability analysis,” IEEE Transactions on Robotics , vol. 40, pp. 4235–4251, 2024

    work page 2024

  6. [6]

    A general hamilton- jacobi framework for non-linear state-constrained control problems,

    A. Altarovici, O. Bokanowski, and H. Zidani, “A general hamilton- jacobi framework for non-linear state-constrained control problems,” ESAIM: Control, Optimisation and Calculus of Variations , vol. 19, no. 2, pp. 337–357, 2013

    work page 2013

  7. [7]

    Cooptimizing safety and performance with a control-constrained formulation,

    H. Wang, A. Dhande, and S. Bansal, “Cooptimizing safety and performance with a control-constrained formulation,” IEEE Control Systems Letters, vol. 8, pp. 2739–2744, 2024

    work page 2024

  8. [8]

    Safe and Performant Deployment of Autonomous Systems via Model Predictive Control and Hamilton-Jacobi Reachability Analysis

    H. Wang, A. Jordana, L. Righetti, and S. Bansal, “Safe and performant deployment of autonomous systems via model predictive control and hamilton-jacobi reachability analysis,” 2025. [Online]. Available: https://arxiv.org/abs/2506.23346

    work page Pith review arXiv 2025

Show all 27 references
  1. [9]

    Learning to be safe: Deep rl with a safety critic,

    K. Srinivasan, B. Eysenbach, S. Ha, J. Tan, and C. Finn, “Learning to be safe: Deep rl with a safety critic,” 2020. [Online]. Available: https://arxiv.org/abs/2010.14603

    2020 arXiv

  2. [10]

    Model-predictive control via cross-entropy and gradient-based optimization,

    H. Bharadhwaj, K. Xie, and F. Shkurti, “Model-predictive control via cross-entropy and gradient-based optimization,” in Proceedings of the 2nd Conference on Learning for Dynamics and Control , ser. Proceedings of Machine Learning Research, A. M. Bayen, A. Jadbabaie, G. Pappas,...

    2020

  3. [11]

    The gradient-based nonlinear model predictive control software grampc,

    B. K ¨apernick and K. Graichen, “The gradient-based nonlinear model predictive control software grampc,” in 2014 European Control Con- ference (ECC), 2014, pp. 1170–1175

    2014

  4. [12]

    Real-time projected gradient-based nonlinear model predictive control with an application to anesthesia control,

    S. Hall, L. Ortmann, M. Picallo, and F. D ¨orfler, “Real-time projected gradient-based nonlinear model predictive control with an application to anesthesia control,” in 2022 IEEE 61st Conference on Decision and Control (CDC), 2022, pp. 6193–6198

    2022

  5. [13]

    Information-theoretic model predictive control: Theory and applica- tions to autonomous driving,

    G. Williams, P. Drews, B. Goldfain, J. M. Rehg, and E. A. Theodorou, “Information-theoretic model predictive control: Theory and applica- tions to autonomous driving,” IEEE Transactions on Robotics, vol. 34, no. 6, pp. 1603–1622, 2018

    2018

  6. [14]

    Multi-agent path integral control for interaction-aware motion planning in urban canals,

    L. Streichenberg, E. Trevisan, J. J. Chung, R. Siegwart, and J. Alonso- Mora, “Multi-agent path integral control for interaction-aware motion planning in urban canals,” in 2023 IEEE International Conference on Robotics and Automation (ICRA) , 2023, pp. 1379–1385

    2023

  7. [15]

    Semi-supervised safe visuomotor policy synthesis using barrier certificates,

    M. Tayal, A. Singh, P. Jagtap, and S. Kolathaya, “Semi-supervised safe visuomotor policy synthesis using barrier certificates,” 2025 IEEE 64th Conference on Decision and Control (CDC) , 2025

    2025

  8. [16]

    Exact imposition of safety boundary conditions in neural reachable tubes,

    A. Singh, Z. Feng, and S. Bansal, “Exact imposition of safety boundary conditions in neural reachable tubes,” in 2025 IEEE International Conference on Robotics and Automation (ICRA) , 2025. [Online]. Available: https://arxiv.org/abs/2404.00814

    2025 arXiv

  9. [17]

    A physics- informed machine learning framework for safe and optimal control of autonomous systems,

    M. Tayal, A. Singh, S. Kolathaya, and S. Bansal, “A physics- informed machine learning framework for safe and optimal control of autonomous systems,” in Forty-second International Conference on Machine Learning , 2025. [Online]. Available: https://openreview.net/forum?id=SrfwiloGQF

    2025

  10. [18]

    Formal synthesis of stochastic systems via control barrier certificates,

    P. Jagtap, S. Soudjani, and M. Zamani, “Formal synthesis of stochastic systems via control barrier certificates,” IEEE Transactions on Auto- matic Control, vol. 66, no. 7, pp. 3097–3110, 2020

    2020

  11. [19]

    Updating quasi-newton matrices with limited storage,

    J. Nocedal, “Updating quasi-newton matrices with limited storage,” Mathematics of Computation , vol. 35, no. 151, pp. 773–782, 1980. [Online]. Available: http://www.jstor.org/stable/2006193

    1980

  12. [20]

    Control barrier function based quadratic programs with application to adaptive cruise control,

    A. D. Ames, J. W. Grizzle, and P. Tabuada, “Control barrier function based quadratic programs with application to adaptive cruise control,” in 53rd IEEE Conference on Decision and Control . IEEE, 2014, pp. 6271–6278

    2014

  13. [21]

    Control barrier functions: Theory and applications,

    A. D. Ames, S. Coogan, M. Egerstedt, G. Notomista, K. Sreenath, and P. Tabuada, “Control barrier functions: Theory and applications,” in 18th European control conference (ECC) . IEEE, 2019, pp. 3420– 3431

    2019

  14. [22]

    Guaranteed-safe mppi through composite control barrier functions for efficient sampling in multi-constrained robotic systems,

    P. Rabiee and J. B. Hoagg, “Guaranteed-safe mppi through composite control barrier functions for efficient sampling in multi-constrained robotic systems,” 2024. [Online]. Available: https://arxiv.org/abs/2410.02154

    2024 arXiv

  15. [23]

    High-order control barrier functions,

    W. Xiao and C. Belta, “High-order control barrier functions,” IEEE Transactions on Automatic Control , vol. 67, no. 7, pp. 3655–3662, 2022

    2022

  16. [24]

    Control barrier functions for systems with high relative degree,

    ——, “Control barrier functions for systems with high relative degree,” in 2019 IEEE 58th Conference on Decision and Control (CDC) , 2019, pp. 474–479

    2019

  17. [25]

    Safety-critical control of a planar quadrotor,

    G. Wu and K. Sreenath, “Safety-critical control of a planar quadrotor,” in 2016 American Control Conference (ACC) , 2016, pp. 2252–2258

    2016

  18. [26]

    Learning a formally verified control barrier function in stochastic environment,

    M. Tayal, H. Zhang, P. Jagtap, A. Clark, and S. Kolathaya, “Learning a formally verified control barrier function in stochastic environment,” in 2024 IEEE 63rd Conference on Decision and Control (CDC), 2024, pp. 4098–4104

    2024

  19. [27]

    Cp-ncbf: A conformal prediction-based approach to synthesize verified neural control barrier functions,

    M. Tayal, A. Singh, P. Jagtap, and S. Kolathaya, “Cp-ncbf: A conformal prediction-based approach to synthesize verified neural control barrier functions,” 2025. [Online]. Available: https://arxiv.org/abs/2503.17395

    2025 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.