Captivity-Escape Games as a Means for Safety in Online Motion Generation

Christopher Bohn; Manuel Hess; S\"oren Hohmann

REVIEW 3 major objections 4 minor 20 references

Captivity-Escape Games as a Means for Safety in Online Motion Generation

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper proves that a reversed differential game—the captivity-escape game—produces a tracking error bound whose worst case is exactly the prescribed safety margin, plus a minimal boundary safety controller.

desk verdict Genuinely new game-theoretic framing with a strong worked example, but the general guarantee that zeta equals alpha is an unproven existence hypothesis, and the authors concede as much in Section VI.E. read the letter →

arxiv 2506.01399 v3 pith:W2DZ2MOT submitted 2025-06-02 eess.SY cs.ROcs.SY

Christopher Bohn , Manuel Hess , Sören Hohmann This is my paper

classification eess.SYcs.ROcs.SY MSC 91A2349N70

keywords captivity-escapegamedifferentialofkindsafetymargintrackingerrorboundworst-casemotiongenerationhomicidalchauffeurcontroller

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 establishes that a safety margin in online motion generation can be used as the input, rather than the output, of a differential game. The authors invert the standard pursuit-evasion setup: an evader modeled by the low-fidelity planner starts inside a captivity set, and the chaser modeled by the high-fidelity tracker tries to keep it there forever. The winning region of this captivity-escape game is an invariant tracking-error bound, and when that region touches the boundary of the captivity set the worst-case tracking error equals the prescribed safety margin exactly. The same construction yields a boundary feedback strategy that acts as a minimal-intervention safety controller. In the homicidal-chauffeur example, solving the game takes about 18 seconds or less, versus roughly 3.13e5 seconds for the FaSTrack state-of-the-art computation, while reproducing the same error bound.

What carries the argument

The captivity-escape game of kind: Player PL, the low-fidelity planning model, starts inside the captivity set Lambda and tries to escape, while Player PH, the high-fidelity tracking model, tries to retain it forever. The captivity zone V+, the set of states from which PH can guarantee eternal captivity, is robustly positively invariant and therefore constitutes the tracking error bound B. Its boundary is constructed from the inward-facing part of the boundary of Lambda, states that PH can defend against immediate escape, and semipermeable surfaces that emanate from the boundary of that inward-facing part and are integrated in retrograde time through an adjoint equation. Nonleaking intersections of these surfaces form a closed barrier K, and the stage at which the surfaces intersect determines the planning performance that makes the worst-case tracking error equal to the margin.

What would settle it

Take the paper's homicidal-chauffeur dynamics and compute the closed barrier for a much smaller safety margin, for instance alpha = 0.05 m with omega_h = 2*pi rad/s and v_h = 1 m/s, by solving the intersection condition. If no positive planning speed v_l exists, or the resulting barrier lies strictly inside the captivity set boundary, then the claimed equality between worst-case tracking error and margin fails for that case.

Watch

Extended reading notes

Core claim

The paper's central discovery is that the captivity zone of a captivity-escape game of kind is exactly the object needed for safe motion generation: it is a robust positively invariant tracking error bound, its size is an upper bound on the worst-case tracking error, and if the captivity zone touches the boundary of the captivity set, the worst-case error is exactly the safety margin. A boundary strategy built from the optimal captivity strategies keeps any planned trajectory inside the tracking error bound for any planner input, and acts only when the state reaches the boundary. The paper solves the game of kind analytically for the homicidal-chauffeur relative dynamics: the semipermeable surfaces emanating from the boundary of the inward-facing part meet at a point that yields a planning speed of about 0.10 m/s for a margin of 0.25 m, reproducing the FaSTrack tracking error bound while avoiding grid-based reachability. The method therefore addresses the paper's objective of directly computing planning performance from a given safety margin, and as a byproduct solves the original safety-margin computation and the safety-controller design.

Load-bearing premise

The headline equality between worst-case tracking error and safety margin holds only if the controller's captivity zone reaches the boundary of the allowed tracking-error set; nothing in the game formulation guarantees that contact for every system pair and every margin.

Editorial extensions

If this is right

Given a prescribed safety margin and under Assumptions 4 and 5, the captivity zone is a tracking error bound with worst-case error exactly equal to the margin, so the safety margin introduces no additional conservatism.
The boundary strategy from Theorem 3 guarantees safe tracking for any planning input and only intervenes on the boundary of the tracking error bound, so it can be paired with any offline or online planner.
The paper's Objective O.1 is solved directly: the planning performance parameter is tuned by solving the barrier conditions, rather than by iteratively recomputing a safety margin.
The same equations solve the original objective O.0, and in the worked example the computation time is about 18 seconds, or under 10 milliseconds in a fast numerical implementation, compared with roughly 3.13e5 seconds for FaSTrack at comparable accuracy.
The method complements existing safe motion generation methods by supplying an accurate, low-cost way to adapt the planning model to a given environment and safety margin.

Reading between the lines

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

The equality between worst-case tracking error and safety margin is a structural property, not a computational accident; for non-homothetic system pairs the likely failure mode is a captivity zone that lies strictly inside the prescribed margin, which would show up as a worst-case error smaller than the margin rather than as a safety violation.
A natural next step is to use the same game to tune parameters beyond scalar speed, such as input limits, quantization, or horizon length, since the formulation already allows a general planning-performance parameter vector.
If the observed agreement with FaSTrack's tracking error bound is proved in general, the captivity-escape construction would provide a certificate for the maximal invariant set without solving the full Hamilton-Jacobi-Isaacs equation, which would change the practical cost of safety verification.
The sub-10-millisecond numerical evaluation time suggests an online scheme in which the safety margin is renegotiated during a mission as obstacles or vehicle capability change; the paper mentions but does not develop this.

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.

Desk Editor's Note

read the letter

Read this one. The captivity-escape game is a real conceptual inversion: instead of computing a safety margin for a fixed planning model, the authors parametrize the planning model and solve for the performance that exactly saturates a given margin. That directly addresses Objective O.1, which the cited FaSTrack, SOS, and primitive-based methods do not. The core theorems are simple and correct given their assumptions, and the writing is clear about what follows from what. The homicidal-chauffeur example is where the paper earns its keep: solving (35) gives v_l approximately 0.10 m/s for alpha equals 0.25 m, and the resulting TEB matches FaSTrack's at about 18 seconds of CPU time versus 3.13e5 seconds on the same machine. That is an orders-of-magnitude speedup, and the numerical-accuracy claim is believable for this analytical example. The soft spots are real but not fatal. The general guarantee zeta equals alpha rests on Assumptions 4 through 6 and on the existence of a nonleaking closed barrier K. No theorem proves such a K exists for arbitrary f, U_l, U_h, and alpha. The stress-test note is right: Corollary 2 needs V+ intersect partial Lambda nonempty, and the constructive method needs the semipermeable surfaces to meet properly in a closed barrier. If the true maximal invariant set sits strictly inside Lambda, the computed WTE is below alpha and O.1 is infeasible for that margin. The paper itself concedes in Section VI.E that the coincidence with FaSTrack is a hypothesis with a formal proof still under development. That is not a numerical-accuracy issue; it is the load-bearing existence premise. Also, the scalar planning-performance parametrization in Assumptions 2 and 3 is restrictive, and no code or data are shipped, so the numerical claims are not independently reproducible as-is. Proportionately, this is a good paper with a missing general theorem. The authors are honest about the gap, and the example is explicit and checked against an external benchmark. The citation pattern is appropriate. A serious referee should ask for either a general existence theorem or a precise characterization of the system pairs for which the barrier exists, plus code or data for the numerical example. Send it to peer review. The core idea deserves referee time, and the revision needs to address the existence gap head-on.

Referee Report

3 major / 4 minor

Summary. The paper introduces a captivity-escape zero-sum differential game in which the high-fidelity tracking model (PH) attempts to keep the low-fidelity planning model (PL) inside a captivity set Λ, while PL attempts to escape. It defines the captivity zone V+ as the set of states from which PH can retain eternal captivity, proves that V+ forms a tracking error bound (Theorem 1), shows ζ ≤ β (Theorem 2), and constructs a minimal-intervention safety controller (Theorem 3). Under Assumptions 4 and 5, Corollary 2 asserts ζ = β = α. The constructive method of Section V builds a compact invariant region C from the inward-facing part of ∂Λ and semipermeable surfaces integrated retrograde from the boundary of the inward-facing part. The numerical example (homicidal chauffeur and pedestrian) solves equation (35) to obtain v_l ≈ 0.10 m/s for α = 0.25 m, reproducing the FaSTrack tracking error bound at orders-of-magnitude lower computational cost.

Significance. If the existence of the closed barrier can be established or the claims are appropriately scoped, the paper makes a valuable contribution: it addresses Objective O.1, which is genuinely unaddressed in the prior literature, and it demonstrates a fast, accurate analytical construction for the homicidal-chauffeur example. The paper's strengths include a clean game formulation, an explicit and falsifiable benchmark against FaSTrack, full reporting of computation times, and an unusually candid discussion of the limits of the general equivalence in Section VI.E. The main obstacle is structural: the central equality ζ = α depends on Assumptions 5 and 6, which are neither proven nor derived from primitive conditions.

major comments (3)

[Section V-A.3 and Assumption 6] The paper's central guarantee (Corollary 2, ζ = α) is not established in the general case. The constructive method builds C from the IP and semipermeable surfaces L integrated retrograde from the BIP via (13)-(16); the safety property requires these surfaces to meet in a nonleaking closed barrier K, but no theorem proves that such a K exists for arbitrary f^ι, U_l, U_h, and α. Assumption 6 merely postulates that ξ_L contains a part ξ_K, and Remark 4 asserts the existence of ν(β) for a 'proper pair' without proof. Section VI.E explicitly concedes that the coincidence with FaSTrack is 'a hypothesis' with 'a formal proof still under development.' If the true maximal invariant set lies strictly in the interior of Λ, then ζ < α and O.1 is infeasible for that margin. The manuscript should either prove existence under stated conditions, or restrict the main theorem to a class of systems where the barrier construction is guaranteed, or downgrade the general claim to a conditional one and state the restriction explicitly in the abstract.
[Section V-B, equations (18)-(22)] The optimization problem that defines ν(β) is not shown to be feasible or well-posed. The paper states that (18), (20), and (21) 'can only be fulfilled for a certain range' of β and ϑ, but it gives no conditions on f^ι, U_l, U_h that guarantee a nonempty feasible set for the offline optimization. If no barrier exists, the optimization is silent and C is undefined, so the claimed online adaption of the planning performance does not follow. Proposition 1 and Proposition 2 provide only local conditions at the BIP; they do not ensure that the retrograde trajectories meet in a closed barrier. A rigorous statement of the assumptions under which ν(β) exists, together with a proof or a counterexample, is needed before the method can be claimed to solve O.1 in general.
[Section VI.E] The numerical validation does not establish that the method computes exactly the FaSTrack TEB. For v_l = 0.50 m/s there is a visible discrepancy (Fig. 3c), and the equal-TEB statement is explicitly presented as a hypothesis with a formal proof 'still under development.' The paper should distinguish the demonstrated contributions (analytical construction, orders-of-magnitude speedup, numerical accuracy for the example) from the conjecture that the method matches the least conservative existing method in general. As written, the discussion and conclusion lean on this unproven equivalence.

minor comments (4)

[Section VI-A, equations (24)-(27)] The reduction from the three-dimensional relative state to (x1, x2) silently discards φ_h with the remark that the planning model 'fully controls' the relative orientation. A few sentences justifying why φ_h becomes obsolete would help the reader follow the derivation.
[Section VI-B, after equation (33)] The statement that 'β/ρ is neglected in (33)' is not self-evident, because ρ is not defined as a normalization constant in the text. Please clarify how the direction of ψ_x is unaffected by this term.
[References] Reference [19] contains a typo: 'Wrigth' should be 'Wright.'
[Section III-B and Definition 12] The notation for the game-of-kind objective and the optimal strategies is inconsistent in places (e.g., J_k versus Jk, and the use of γ• versus γ⋆). A unified notation would improve readability.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the ζ=α claim is a conditional corollary of explicit assumptions, and the numerical agreement with FaSTrack is an external benchmark, not a fitted target.

full rationale

The derivation chain is self-contained. The bound ζ ≤ β follows directly from the set inclusion V⁺ ⊆ Λ (Theorem 2), and ζ = β = α is obtained only under the explicitly stated Assumption 5 that V⁺ touches ∂Λ. This is a sufficient geometric condition, not a conclusion fed back into the definition of ζ or α. The construction of C from the inward-facing part and semipermeable surfaces follows Isaacs' classical theory rather than any self-citation; none of the paper's load-bearing references are authored by Bohn, Hess, or Hohmann. The numerical value v_l ≈ 0.10 m/s is obtained by solving Eq. (35) with β = α = 0.25 m, not by fitting to the FaSTrack output. The matching TEB for v_l = 0.10 m/s is therefore an external benchmark comparison, and the paper explicitly labels the general equivalence with FaSTrack as a hypothesis with a proof still under development. The connectedness ansatz ν(β) = [0 β]ᵀ is a stated design constraint used for comparison, not a parameter fitted to data. The unimplemented general existence of a nonleaking closed barrier K is a correctness or assumption-risk issue, not a circularity: it does not make the derived ζ=α statement equivalent to its inputs by construction. No circular step meeting the evidentiary bar was found.

Assumptions & free parameters 0 free parameters · 7 assumptions · 0 invented entities

The central derivation leans on classical Isaacs barrier theory for the construction of the invariant set C, plus several assumptions specific to this paper: scalar performance parametrization, nonempty and compact V+, boundary contact V+ intersect partial Lambda, and the barrier-part condition. These are stated but not proven for general systems and are the main source of correctness risk.

assumptions (7)

domain assumption Assumption 1: the high-fidelity model accurately represents the real-world system dynamics.
Safety guarantees are stated with respect to the high-fidelity model, so model error transfers to the real system. This is standard in model-based safety and not proven by the paper.
ad hoc to paper Assumptions 2-3: the low-fidelity sets X_l, U_l and the parameter vector iota_l are fully determined by a single scalar planning performance vartheta through a monotone mapping o(vartheta).
This restricts O.1 to scalar performance trade-offs and no general construction of the mapping o is given; it is introduced specifically to make O.1 tractable.
domain assumption Assumption 4: the captivity zone V+ is nonempty and compact.
Required for V+ to qualify as a tracking error bound with a well-defined WTE. Not proven for all system pairs.
ad hoc to paper Assumption 5: V+ intersects the boundary of the captivity set, V+ intersect partial Lambda is nonempty.
Needed for Corollary 2 so that zeta = beta = alpha. The paper says the construction makes this automatic, but no general proof is given.
ad hoc to paper Assumption 6: each semipermeable trajectory xi_L contains a barrier part xi_K.
Used in Proposition 2 and Corollary 5 to ensure that trajectories passing the BIP enter C rather than the exterior of C.
standard math Isaacs' semipermeable-surface and closed-barrier theory is valid and directly applicable.
The construction of C subset V+ depends on this classical theory, cited from [16, Chap. 8], [14], and [17]; the paper does not re-derive it.
standard math Caratheodory and Filippov solution frameworks for open-loop and closed-loop discontinuous dynamics.
Remark 2 cites [15] and [7] to justify existence and uniqueness of trajectories under switching safety controllers.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Captivity-Escape Games as a Means for Safety in Online Motion Generation." pith.science (2026). https://pith.science/paper/W2DZ2MOT

@misc{pith2026250601399,
  author       = {Pith},
  title        = {Pith review of: Captivity-Escape Games as a Means for Safety in Online Motion Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/W2DZ2MOT}},
  note         = {Machine review of arXiv:2506.01399}
}

read the original abstract

This paper addresses conservatism, limited numerical accuracy, and high computational effort in existing methods ensuring safety by design in online model-based motion generation. The presented method employs a novel captivity-escape zero-sum differential game to adapt the planning model's performance so that resulting reference trajectories are trackable within a prescribed safety margin by a jointly synthesized safety controller. A numerical example demonstrates orders-of-magnitude faster computation and improved numerical accuracy compared to the state of the art.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

20 extracted references · 20 canonical work pages

[1]

Fastrack: A modular framework for fast and guaranteed safe motion planning,

S. L. Herbert et al., “Fastrack: A modular framework for fast and guaranteed safe motion planning,” in2017 IEEE 56th Conf. Decis. Control, 2017, pp. 1517–1522

work page 2017
[2]

Fastrack: A modular framework for real-time motion planning and guaranteed safe tracking,

M. Chen et al., “Fastrack: A modular framework for real-time motion planning and guaranteed safe tracking,”IEEE Trans. Autom. Control, vol. 66, no. 12, pp. 5861–5876, 2021

work page 2021
[3]

Robust tracking with model mismatch for fast and safe planning: An sos optimization approach,

S. Singh, M. Chen, S. L. Herbert, C. J. Tomlin, and M. Pavone, “Robust tracking with model mismatch for fast and safe planning: An sos optimization approach,” inAlgorithmic Found. Robot. XIII, Cham: Springer, 2020, pp. 545–564

work page 2020
[4]

Bridging the gap between safety and real-time performance in receding-horizon trajectory design for mobile robots,

S. Kousik, S. Vaskov, F. Bu, M. Johnson-R., and R. Vasudevan, “Bridging the gap between safety and real-time performance in receding-horizon trajectory design for mobile robots,”Int. J. Robot. Res., vol. 39, no. 12, pp. 1419–1469, 2020

work page 2020
[5]

Formal safety net control using backward reachability anal- ysis,

B. Sch ¨urmann, M. Klischat, N. Kochdumper, and M. Althoff, “Formal safety net control using backward reachability anal- ysis,”IEEE Trans. Autom. Control, vol. 67, no. 11, pp. 5698– 5713, 2022

work page 2022
[6]

A general safety framework for learning- based control in uncertain robotic systems,

J. F. Fisac et al., “A general safety framework for learning- based control in uncertain robotic systems,”IEEE Trans. Autom. Control, vol. 64, no. 7, pp. 2737–2752, 2019

work page 2019
[7]

A time-dependent hamilton-jacobi formulation of reachable sets for continuous dynamic games,

I. Mitchell, A. Bayen, and C. Tomlin, “A time-dependent hamilton-jacobi formulation of reachable sets for continuous dynamic games,”IEEE Trans. Autom. Control, vol. 50, no. 7, pp. 947–957, 2005

work page 2005
[8]

E. A. Coddington and N. Levinson,Theory of Ordinary Differential Equations. Melbourne, FL, USA: Krieger, 1984

work page 1984

Show all 20 references

[9]

A Computationally Efficient Robust Model Predictive Control Framework for Uncertain Nonlinear Systems,

J. K ¨ohler, R. Soloperto, M. A. M ¨uller, and F. Allg ¨ower, “A Computationally Efficient Robust Model Predictive Control Framework for Uncertain Nonlinear Systems,”IEEE Trans. Autom. Control, vol. 66, no. 2, pp. 794–801, 2021

2021
[10]

Planning, fast and slow: A framework for adaptive real-time safe trajectory planning,

D. Fridovich-Keil, S. L. Herbert, J. F. Fisac, S. Deglurkar, and C. J. Tomlin, “Planning, fast and slow: A framework for adaptive real-time safe trajectory planning,” in2018 IEEE Int. Conf. Robot. Automat., 2018, pp. 387–394

2018
[11]

Pa-fastrack: Planner- aware real-time guaranteed safe planning,

A. Sahraeekhanghah and M. Chen, “Pa-fastrack: Planner- aware real-time guaranteed safe planning,” in2021 60th IEEE Conf. Decis. Control, 2021, pp. 2129–2136

2021
[12]

Quarteroni, R

A. Quarteroni, R. Sacco, and F. Saleri,Numerical Mathemat- ics. Berlin Heidelberg: Springer, 2010

2010
[13]

Differential games and representation formulas for hamilton-jacobi-isaacs equations,

L. C. Evans and P. E. Souganidis, “Differential games and representation formulas for hamilton-jacobi-isaacs equations,” Indiana Univ. Math. J., vol. 33, no. 5, pp. 773–797, 1984

1984
[14]

Pursuit-Evasion Games,

V . Patsko, S. Kumkov, and V . Turova, “Pursuit-Evasion Games,” inHandbook Dyn. Game Theory, Cham: Springer, 2018

2018
[15]

Discontinuous dynamical systems,

J. Cortes, “Discontinuous dynamical systems,”IEEE Control Syst. Mag., vol. 28, no. 3, pp. 36–73, 2008

2008
[16]

Isaacs,Differential Games, New ed Edition

R. Isaacs,Differential Games, New ed Edition. Dover Publi- cations Inc., 1999

1999
[17]

The Game of Two Identical Cars: An Analytical Description of the Barrier,

M. Buzikov and A. Galyaev, “The Game of Two Identical Cars: An Analytical Description of the Barrier,”J. Optim. Theory Appl., vol. 198, no. 3, pp. 988–1018, 2023

2023
[18]

Lewin,Differential Games

J. Lewin,Differential Games. London: Springer, 1994

1994
[19]

Nocedal and S

J. Nocedal and S. J. Wrigth,Numerical Optimization(Springer Ser. Operations Res. Financial Eng.). New York: Springer, 2006

2006
[20]

Application of level set methods to control and reachability problems in continuous and hybrid system,

I. M. Mitchell, “Application of level set methods to control and reachability problems in continuous and hybrid system,” Ph.D. dissertation, Stanford Univ., 2002

2002

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Fastrack: A modular framework for fast and guaranteed safe motion planning,

S. L. Herbert et al., “Fastrack: A modular framework for fast and guaranteed safe motion planning,” in2017 IEEE 56th Conf. Decis. Control, 2017, pp. 1517–1522

work page 2017

[2] [2]

Fastrack: A modular framework for real-time motion planning and guaranteed safe tracking,

M. Chen et al., “Fastrack: A modular framework for real-time motion planning and guaranteed safe tracking,”IEEE Trans. Autom. Control, vol. 66, no. 12, pp. 5861–5876, 2021

work page 2021

[3] [3]

Robust tracking with model mismatch for fast and safe planning: An sos optimization approach,

S. Singh, M. Chen, S. L. Herbert, C. J. Tomlin, and M. Pavone, “Robust tracking with model mismatch for fast and safe planning: An sos optimization approach,” inAlgorithmic Found. Robot. XIII, Cham: Springer, 2020, pp. 545–564

work page 2020

[4] [4]

Bridging the gap between safety and real-time performance in receding-horizon trajectory design for mobile robots,

S. Kousik, S. Vaskov, F. Bu, M. Johnson-R., and R. Vasudevan, “Bridging the gap between safety and real-time performance in receding-horizon trajectory design for mobile robots,”Int. J. Robot. Res., vol. 39, no. 12, pp. 1419–1469, 2020

work page 2020

[5] [5]

Formal safety net control using backward reachability anal- ysis,

B. Sch ¨urmann, M. Klischat, N. Kochdumper, and M. Althoff, “Formal safety net control using backward reachability anal- ysis,”IEEE Trans. Autom. Control, vol. 67, no. 11, pp. 5698– 5713, 2022

work page 2022

[6] [6]

A general safety framework for learning- based control in uncertain robotic systems,

J. F. Fisac et al., “A general safety framework for learning- based control in uncertain robotic systems,”IEEE Trans. Autom. Control, vol. 64, no. 7, pp. 2737–2752, 2019

work page 2019

[7] [7]

A time-dependent hamilton-jacobi formulation of reachable sets for continuous dynamic games,

I. Mitchell, A. Bayen, and C. Tomlin, “A time-dependent hamilton-jacobi formulation of reachable sets for continuous dynamic games,”IEEE Trans. Autom. Control, vol. 50, no. 7, pp. 947–957, 2005

work page 2005

[8] [8]

E. A. Coddington and N. Levinson,Theory of Ordinary Differential Equations. Melbourne, FL, USA: Krieger, 1984

work page 1984