Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

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

T0 review · 4 major / 6 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 Inserting the Hamilton-Jacobi safety value function as a final-time constraint in MPC makes the controller recursively feasible: from any initially safe state, every future planning step has a feasible plan, so long-horizon safety holds…

desk verdict Nice combination of HJ reachability and MPC, but the recursive-feasibility guarantee has a discrete-continuous gap and the novelty over prior work is thin. read the letter →

arxiv 2506.23346 v1 pith:LFKVORAQ submitted 2025-06-29 cs.RO cs.SYeess.SY

Hao Wang , Armand Jordana , Ludovic Righetti , Somil Bansal This is my paper
classification cs.ROcs.SYeess.SY
keywords safetyvaluefunctionmodelpredictivecontrolHamilton-Jacobireachabilityrecursivefeasibilityautonomoussystemsconstraints6-DoFmanipulatorDubinscar
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

This paper proposes a way to make model predictive control (MPC) safe over long time horizons without sacrificing task performance: add a terminal constraint requiring the final state of each short-horizon plan to lie in the super-zero level set of the Hamilton-Jacobi safety value function, $\{x : V_s(x) \geq 0\}$. The paper argues, via Lemma 1, that this terminal constraint makes the MPC recursively feasible: if the initial state is in the maximal set of recursively feasible states, a feasible solution exists at every future step, so the closed-loop system never gets forced into a safety violation. The framework is demonstrated on a 4D Dubins car and a 6-DoF Kuka manipulator, where it substantially improves safety-constraint satisfaction over baseline MPC without a formal safety term. When the safety value function is computed exactly (as for the Dubins car), the guarantee is formal; when it is learned (as for the Kuka), the safety improvement is empirical, since the learned approximation is not exact.

What carries the argument

The safety value function $V_s(x)$ — the value of the safety optimal control problem that maximizes the minimal safety margin $\min_\tau l(\xi(\tau))$ over all future times — is the machine that carries the argument. Its super-zero level set $\{x : V_s(x) \ge 0\}$ is the maximal set of recursively feasible states: every state in it can keep the system safe forever, and no state outside it can. Inserting $V_s(x(j+h)) \ge 0$ as the terminal constraint of the MPC (Problem 5) means every planning horizon ends inside a state from which safety can be maintained indefinitely; Lemma 1 then propagates feasibility from one MPC step to the next. The same object also makes the method scalable, because computing $V_s$ is a separate offline step that can be carried out by grid-based level-set methods in low dimensions or by learned approximators in high dimensions.

What would settle it

For the 4D Dubins car, where the safety value function is computed exactly on a grid, run the closed-loop Safety Value MPC from every grid state in the super-zero level set for the full 2-second task horizon; finding any state from which the trajectory violates the obstacle constraint, or any step at which Problem 5 becomes infeasible, would contradict Lemma 1.

Watch

Extended reading notes

Core claim

The central claim is that a terminal constraint built from the converged Hamilton-Jacobi safety value function $V_s(x)=\sup_u \min_{\tau\in[t,\infty)} l(\xi^u_{x,t}(\tau))$ turns any finite-horizon MPC into a recursively feasible controller. Because the super-zero level set of $V_s$ is exactly the maximal set of states from which the system can respect the safety constraint forever, a plan that ends anywhere inside that set can always be extended to another feasible plan at the next MPC step. Lemma 1 formalizes this: starting from a recursively feasible state, the Safety Value MPC (Problem 5) is feasible at every step $k$, under the assumption of an exact value function and successful solver returns. The significance is that the safety value function encodes all long-horizon safety reasoning offline, so online MPC only needs a short horizon to achieve persistent constraint satisfaction, and the same value function works for any task cost.

Load-bearing premise

The recursive feasibility guarantee assumes the exact continuous-time safety value function characterizes all recursively feasible states and that every MPC solve returns a feasible solution ending in that set; for the 12D manipulator, the learned value function is not exact, so the formal guarantee does not apply there.

Editorial extensions

If this is right

  • If the safety value function is exact, recursive feasibility is guaranteed for any planning horizon, so operators can choose shorter horizons for computational speed without sacrificing long-horizon safety.
  • The safety value function is independent of the task cost, so one offline reachability computation supports multiple goals and performance objectives; the MPC simply optimizes the task cost subject to the same terminal safety constraint.
  • The terminal constraint is compatible with high-level planners that emit cost functions (including LLM/VLM planners): the safety guarantee is preserved as long as the MPC respects the pre-defined constraints, regardless of the planner's incentives.
  • In high-dimensional settings where the value function is learned, the framework still improves safety-constraint satisfaction empirically, as shown by the 6-DoF Kuka experiments (11/15 success vs 3/15 baseline).
  • The method's task performance is slightly more conservative than unconstrained MPC, but the simulations show the conservatism is small and shrinks as the planning horizon grows.

Reading between the lines

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

  • The formal guarantee inherits a caveat: it is a property of the exact value function; for learned approximations, a natural test would be to compare the learned zero level set against the true maximal safe set on a validation grid, which the paper does not do.
  • Because the value function is task-independent, the framework's computational cost can be amortized: a single offline reachability computation can serve an entire family of online tasks, suggesting practical deployment for goal-to-goal navigation or manipulation.
  • The guarantee is discrete-time: only the MPC sample points are certified, not the trajectory in between. A testable extension would be to combine the terminal value constraint with a low-level robust tracking controller to bound inter-sample violation.
  • For high-dimensional robots, if a neural verifier could certify that the learned value function's zero-level set sits inside the true maximal safe set, Lemma 1 would formally extend; the authors point to verification techniques as a future direction.
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

4 major / 6 minor

Summary. The paper proposes an MPC formulation (Problem 5) in which a Hamilton-Jacobi reachability safety value function Vs is imposed as a terminal constraint Vs(x(j+h)) >= 0. The authors argue that this makes the MPC recursively feasible (Lemma 1), because the super-0-level set of Vs is the maximal set of states from which the continuous-time safety constraints can be satisfied indefinitely. The method is demonstrated on a 4D Dubins car with an exact grid-based value function and on a 12D Kuka manipulator with a learned value function, with empirical results showing improved safety over an MPC baseline.

Significance. If the recursive-feasibility guarantee were established at the level of rigor claimed, the construction would be attractive: it uses the maximal recursively feasible set as a terminal constraint instead of a more conservative control-invariant subset, and it separates safety reasoning from the task cost. The Dubins experiments show a clear qualitative and quantitative improvement, and the learned 12D value function indicates a plausible scalability path. The paper also avoids fitting free parameters for the safety constraint, since the value function is obtained by HJ computation or learning rather than tuned against the experiments. However, the formal content is currently limited by the gap between the continuous-time value function and the discrete-time MPC, and by the paper's own admission that the high-dimensional experiment carries no formal guarantee.

major comments (4)
  1. [III.B, Lemma 1 and Prob. 5] The proof transfers a continuous-time invariance property to the discrete-time MPC without a consistency assumption. The safety value function Vs in Eq. (5) is defined through the continuous dynamics (1b), whereas Prob. 5 is solved with the discrete dynamics fd in Eq. (6b); the proof identifies {x | Vs(x) >= 0} with the recursively feasible set for Prob. 5, but this identification is not established and can fail for an approximate discretization such as an Euler or RK scheme. The recursive-feasibility guarantee therefore does not follow for the controller actually implemented unless fd is the exact sampled-data model of f, or unless a quantified discretization-error argument is supplied. Please state such an assumption explicitly and prove the corresponding discrete invariance, or restrict the claim to the exact-sampled case.
  2. [III.A-B, Prob. 5 and Eq. (1c)] Safety is enforced only at the sampling instants in Eq. (6c), while the original constraint (1c) is continuous in time. Lemma 1 does not provide any bound on l between samples. Since the motivating application is obstacle avoidance, a trajectory can enter the failure set between two MPC samples and leave it before the next sample; recursive feasibility alone does not rule this out. An inter-sample safety argument (e.g., Lipschitz or barrier-function estimates, or an inner-approximated continuous-time invariant set) is needed before the closed-loop system can be said to respect the safety constraints.
  3. [IV.B and V] For the 12D Kuka experiment the safety value function is learned, and Section V explicitly states that no formal safety guarantee currently holds. Therefore Lemma 1 is not applicable to the high-dimensional demonstration, and the abstract's unconditional claim that the framework 'guarantees recursive feasibility' is not supported. The formal guarantee should be scoped to settings with an exact or verified value function, and the 11/15 success in the Kuka study should be presented as empirical evidence rather than as a guaranteed outcome.
  4. [III.B, Lemma 1] The lemma statement omits two assumptions that the proof uses: the absence of disturbances and the ability of the solver to return a feasible solution at every invocation. The phrase 'assuming the absence of disturbances' appears only inside the proof. Because the implementation uses Crocoddyl, a local SQP solver, the guarantee is conditional on the solver's success; these assumptions should be stated as part of the lemma for the claimed guarantee to be meaningful.
minor comments (6)
  1. [Prob. 5, Eq. (6c)] The constraint index set in Eq. (6c) excludes the terminal index k=j+h, unlike Prob. 3; the inconsistency should be fixed, and it should be noted that for an exact Vs the terminal constraint implies l(x(j+h)) >= 0.
  2. [IV.A] The text says 'the running cost and the terminal cost, defined in Eq. (6b)'; the costs are defined in Eq. (6a), not Eq. (6b).
  3. [Eq. (5)] The superscript d in xi^{u,d}_{x,t} is never defined; it should be introduced, or the notation should be simplified.
  4. [V] The sentence 'our method replies on learning-based methods' contains a typo; 'replies' should be 'relies'.
  5. [Table I] The definition of 'Rollout Success Rate' and the number of trials should be stated explicitly; currently only percentages are given, and it is unclear whether success means safe arrival at the goal or merely absence of constraint violation.
  6. [IV.B] The text says the safety value function is computed by 'a learning-based approach [6, 9]' but does not specify which method is used, what approximation error is observed, or whether any validation or verification was performed; at least the key hyperparameters and validation procedure should be reported.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the recursive-feasibility claim rests on an imported HJ reachability characterization, not on a fit, rename, or definitional identity.

full rationale

The paper's central theoretical claim, Lemma 1, states that Safety Value MPC (Prob. 5) is recursively feasible whenever the initial state lies in {Vs >= 0}. This is not circular: Vs is defined independently by the HJ reachability optimal control problem (Prob. 4, Eq. (5)), and the terminal constraint (6d) is a standard control-invariant terminal set for the MPC. The paper imports from its own prior work [8] the fact that {Vs >= 0} is the maximal set of recursively feasible states; that is a substantive external mathematical theorem with stated dynamics and constraint assumptions, not a fitted parameter or a restatement of the MPC objective. No quantity is fitted to data and then reported as a prediction: the Dubins safety value function is computed offline with LevelSetToolBox/HelperOC, and the Kuka value function is learned, with the paper explicitly conceding in Sec. V that no formal safety guarantee holds for the learned case. The observations about a missing discrete/continuous consistency condition between fd and f, and the questionable inference that all MPC solution states are recursively feasible, are soundness gaps in the proof, not circularities: they do not make the conclusion equivalent to the assumptions by construction. Self-citation of [8] is load-bearing, but it cites a prior peer-reviewed characterization external to the present paper's fitted values, so under the review rules it is independent support and does not raise the circularity score.

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

No numerical constants are fitted to data; the experimental settings (horizons, grid sizes, acceleration bounds) are simulation choices, not calibrated parameters. The central claim rests on the standard HJ viability characterization, the accuracy of a learned value function in the high-dimensional experiment, the global optimality of the MPC solver, and discrete-time constraint sampling. No new entities are postulated.

assumptions (4)
  • standard math The super-0 level set of the converged HJ safety value function equals the set of all recursively feasible states under the continuous-time dynamics.
    Invoked in the proof of Lemma 1 (Sec. III-B) and attributed to [8]; this is a standard viability/reachability characterization, but its regularity assumptions are not checked here.
  • ad hoc to paper The learned safety value function used for the 12D manipulator is an accurate enough approximation of the true safety value function.
    Sec. IV-B uses a learning-based approach [6,9], and Sec. V admits it provides no formal safety guarantees; the 11/15 success claim depends on this approximation.
  • ad hoc to paper Each solve of Prob. 5 by Crocoddyl yields a feasible solution whose terminal state lies in the recursive feasible set.
    Lemma 1 assumes the MPC solution is available and exact; SQP solvers generally return local solutions, so recursive feasibility may fail in practice.
  • domain assumption Constraint satisfaction at discrete MPC sample times implies safety of the continuous-time trajectory.
    Prob. 3 and Prob. 5 enforce l(x(k)) only at discrete indices; the safety value function is computed in continuous time, but inter-sample violations are not excluded.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Safe and Performant Deployment of Autonomous Systems via Model Predictive Control and Hamilton-Jacobi Reachability Analysis." pith.science (2026). https://pith.science/paper/LFKVORAQ

@misc{pith2026250623346,
  author       = {Pith},
  title        = {Pith review of: Safe and Performant Deployment of Autonomous Systems via Model Predictive Control and Hamilton-Jacobi Reachability Analysis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LFKVORAQ}},
  note         = {Machine review of arXiv:2506.23346}
}
read the original abstract

While we have made significant algorithmic developments to enable autonomous systems to perform sophisticated tasks, it remains difficult for them to perform tasks effective and safely. Most existing approaches either fail to provide any safety assurances or substantially compromise task performance for safety. In this work, we develop a framework, based on model predictive control (MPC) and Hamilton-Jacobi (HJ) reachability, to optimize task performance for autonomous systems while respecting the safety constraints. Our framework guarantees recursive feasibility for the MPC controller, and it is scalable to high-dimensional systems. We demonstrate the effectiveness of our framework with two simulation studies using a 4D Dubins Car and a 6 Dof Kuka iiwa manipulator, and the experiments show that our framework significantly improves the safety constraints satisfaction of the systems over the baselines.

Figures

Figures reproduced from arXiv: 2506.23346 by the authors.

Figure 1
Figure 1. Dubins4D obstacle configuration and rollouts for our method and [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. End-effector trajectories of our method and the MPC baseline for [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

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

    eess.SY 2025-07 reject novelty 4.0 of 10

    A two-stage controller that uses L-BFGS gradient-based MPC for performance and a CBF-QP filter for hard safety constraints is demonstrated on simulated unicycle and planar quadrotor navigation.

Reference graph

Works this paper leans on

25 extracted references · 11 canonical work pages · cited by 1 Pith paper

  1. [8]

    On safety and liveness filtering using hamilton-jacobi reachability analysis

    Javier Borquez, Kaustav Chakraborty, Hao Wang, and Somil Bansal. On safety and liveness filtering using hamilton-jacobi reachability analysis. IEEE Transactions on Robotics, 2024

    work page 2024

  2. [25]

    Coop- timizing Safety and Performance With a Control- Constrained Formulation

    Hao Wang, Adityaya Dhande, and Somil Bansal. Coop- timizing Safety and Performance With a Control- Constrained Formulation. IEEE Control Systems Letters, 2024

    work page 2024

  3. [1]

    https://github.com/ HJReachability/helperOC

    helperOC Library, 2019. https://github.com/ HJReachability/helperOC

    work page 2019

  4. [2]

    Gpt-4 technical report

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023

    arXiv 2023

  5. [3]

    Do as i can, not as i say: Grounding language in robotic affordances

    Michael Ahn, Anthony Brohan, Noah Brown, Yevgen Chebotar, Omar Cortes, Byron David, Chelsea Finn, Chuyuan Fu, Keerthana Gopalakrishnan, Karol Haus- man, et al. Do as i can, not as i say: Grounding language in robotic affordances. arXiv preprint arXiv:2204.01691, 2022

    arXiv 2022

  6. [4]

    A general Hamilton-Jacobi framework for non- linear state-constrained control problems

    Albert Altarovici, Olivier Bokanowski, and Hasnaa Zi- dani. A general Hamilton-Jacobi framework for non- linear state-constrained control problems. ESAIM: COCV, 19(2):337–357, 2013. doi: 10.1051/cocv/ 2012011. URL https://doi.org/10.1051/cocv/2012011

    arXiv 2013

  7. [5]

    Control barrier function based quadratic programs for safety critical systems

    Aaron D Ames, Xiangru Xu, Jessy W Grizzle, and Paulo Tabuada. Control barrier function based quadratic programs for safety critical systems. IEEE Transactions on Automatic Control , 62(8):3861–3876, 2016

    work page 2016

  8. [6]

    Deepreach: A deep learning approach to high-dimensional reachability

    Somil Bansal and Claire J Tomlin. Deepreach: A deep learning approach to high-dimensional reachability. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 1817–1824. IEEE, 2021

    2021

Show all 25 references
  1. [7]

    Conservative safety critics for exploration

    Homanga Bharadhwaj, Aviral Kumar, Nicholas Rhine- hart, Sergey Levine, Florian Shkurti, and Animesh Garg. Conservative safety critics for exploration. arXiv preprint arXiv:2010.14497, 2020

    2010 arXiv

  2. [9]

    Bridging model predictive control and deep learning for scalable reachability analysis

    Zeyuan Feng, Le Qiu, and Somil Bansal. Bridging model predictive control and deep learning for scalable reachability analysis. arXiv preprint arXiv:2505.03830 , 2025

    2025 arXiv

  3. [10]

    A survey on hallucination in large language models: Principles, taxon- omy, challenges, and open questions

    Lei Huang, Weijiang Yu, Weitao Ma, Weihong Zhong, Zhangyin Feng, Haotian Wang, Qianglong Chen, Weihua Peng, Xiaocheng Feng, Bing Qin, et al. A survey on hallucination in large language models: Principles, taxon- omy, challenges, and open questions. ACM Transactions on Informat...

    2025

  4. [11]

    π0.5: a vision-language-action model with open-world generalization

    Physical Intelligence, Kevin Black, Noah Brown, James Darpinian, Karan Dhabalia, Danny Driess, Adnan Es- mail, Michael Equi, Chelsea Finn, Niccolo Fusai, et al. π0.5: a vision-language-action model with open-world generalization. arXiv preprint arXiv:2504.16054 , 2025

    2025 arXiv

  5. [12]

    Stagewise implementations of sequential quadratic programming for model-predictive control

    Armand Jordana, S ´ebastien Kleff, Avadesh Meduri, Justin Carpentier, Nicolas Mansard, and Ludovic Righetti. Stagewise implementations of sequential quadratic programming for model-predictive control

  6. [13]

    Position: Llms can’t plan, but can help planning in llm-modulo frame- works

    Subbarao Kambhampati, Karthik Valmeekam, Lin Guan, Mudit Verma, Kaya Stechly, Siddhant Bhambri, Lu- cas Paul Saldyt, and Anil B Murthy. Position: Llms can’t plan, but can help planning in llm-modulo frame- works. In Forty-first International Conference on Ma- chine Learning, 2024

    2024

  7. [14]

    Certifying llm safety against adversarial prompting

    Aounon Kumar, Chirag Agarwal, Suraj Srinivas, Aaron Jiaxun Li, Soheil Feizi, and Himabindu Lakkaraju. Certifying llm safety against adversarial prompting. arXiv preprint arXiv:2309.02705 , 2023

    2023 arXiv

  8. [15]

    Code as policies: Language model programs for em- bodied control

    Jacky Liang, Wenlong Huang, Fei Xia, Peng Xu, Karol Hausman, Brian Ichter, Pete Florence, and Andy Zeng. Code as policies: Language model programs for em- bodied control. In 2023 IEEE International Conference on Robotics and Automation (ICRA) , pages 9493–9500. IEEE, 2023

    2023

  9. [16]

    Generating formal safety assurances for high-dimensional reachability

    Albert Lin and Somil Bansal. Generating formal safety assurances for high-dimensional reachability. In 2023 IEEE International Conference on Robotics and Automa- tion (ICRA), pages 10525–10531. IEEE, 2023

    2023

  10. [17]

    Verification of neural reachable tubes via scenario optimization and conformal prediction

    Albert Lin and Somil Bansal. Verification of neural reachable tubes via scenario optimization and conformal prediction. In 6th Annual Learning for Dynamics & Control Conference, pages 719–731. PMLR, 2024

    2024

  11. [18]

    Crocoddyl: An efficient and versatile framework for multi-contact optimal control

    Carlos Mastalli, Rohan Budhiraja, Wolfgang Merkt, Guilhem Saurel, Bilal Hammoud, Maximilien Naveau, Justin Carpentier, Ludovic Righetti, Sethu Vijayakumar, and Nicolas Mansard. Crocoddyl: An efficient and versatile framework for multi-contact optimal control. In 2020 IEEE Inte...

    2020

  12. [19]

    A toolbox of level set methods

    Ian M Mitchell et al. A toolbox of level set methods. UBC Department of Computer Science Technical Report TR-2007-11, page 31, 2007

    2007

  13. [20]

    Progprompt: Generating situated robot task plans using large language models

    Ishika Singh, Valts Blukis, Arsalan Mousavian, Ankit Goyal, Danfei Xu, Jonathan Tremblay, Dieter Fox, Jesse Thomason, and Animesh Garg. Progprompt: Generating situated robot task plans using large language models. In 2023 IEEE International Conference on Robotics and Automatio...

    2023

  14. [21]

    Learning to be safe: Deep rl with a safety critic

    Krishnan Srinivasan, Benjamin Eysenbach, Sehoon Ha, Jie Tan, and Chelsea Finn. Learning to be safe: Deep rl with a safety critic. arXiv preprint arXiv:2010.14603 , 2020

    2010 arXiv

  15. [22]

    Gemini robotics: Bringing ai into the physical world

    Gemini Robotics Team, Saminda Abeyruwan, Joshua Ainslie, Jean-Baptiste Alayrac, Montserrat Gonzalez Arenas, Travis Armstrong, Ashwin Balakrishna, Robert Baruch, Maria Bauza, Michiel Blokzijl, et al. Gemini robotics: Bringing ai into the physical world. arXiv preprint arXiv:250...

    2025 arXiv

  16. [23]

    Attention is all you need

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems , 30, 2017

    2017

  17. [24]

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

    Kim P Wabersich, Andrew J Taylor, Jason J Choi, Koushil Sreenath, Claire J Tomlin, Aaron D Ames, and Melanie N Zeilinger. Data-driven safety filters: Hamilton-jacobi reachability, control barrier functions, and predictive methods for uncertain systems. IEEE Control Systems Mag...

    2023

Pith tools

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