{"id":"28af44f9-a4d8-43ac-a97c-084b0573cb9a","arxiv_id":"2506.01399","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A captivity-escape differential game computes the planner model's performance needed to keep tracking error inside a safety margin, and in the numerical example it runs orders of magnitude faster than FaSTrack.","lead":"This paper introduces a reversed differential game, called a captivity-escape game, to make safety-guaranteed robot motion planning much faster and numerically more reliable. It lets planners adapt their model's performance online so that reference trajectories stay within a prescribed safety margin.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The ζ=α guarantee is not proven in general: it depends on an unproven existence of a nonleaking closed barrier K (Assumption 5/6), and the paper itself concedes in Section VI.E that the general equivalence is only a hypothesis.","rationale":"The paper's central contribution is a method that solves O.1 and O.2 with an exact margin, and the hinge is Corollary 2: the constructed TEB has WTE exactly equal to α. That statement requires V+ to touch ∂Λ and, in the constructive algorithm, requires the retrograde semipermeable surfaces to form a nonleaking closed barrier K. Neither condition is established by a theorem in the manuscript. The homicidal-chauffeur example verifies one instance analytically, but Section VI.E itself states that the general coincidence with FaSTrack TEBs is a hypothesis under development. This is a real structural gap, because if no barrier exists for some pair and margin, the optimization may return no feasible ϑ, or worse, an apparent solution that is not actually invariant. The reader's weakest_assumption identifies exactly this combination of Assumption 5 and the existence of a nonleaking K. I therefore agree with the CONDITIONAL verdict: the theoretical and numerical results are coherent as far as they go, but the generality of the central guarantee is unproven. A concrete second-benchmark test with a full value-function comparison would settle whether the concern lands in practice.","tokens_in":15198,"tokens_out":17154,"duration_ms":215059,"concrete_test":"On a second benchmark pair from the FaSTrack literature (e.g., a single-integrator planner with speed bound v_l and a unicycle tracker with speed v_h and yaw-rate bound ω_h), fix α and compute the proposed C by integrating (13)-(15) backward from the BIP and solving (18)-(22) for ϑ. Then compute the true game value V and true V+ on a fine grid with helperOC and check (1) whether V+∩∂Λ is nonempty and (2) whether the constructed C is contained in V+ and is invariant under the boundary controller (23) by forward simulation with adversarial u_l. If either check fails for some α reachable by the optimization, then the paper's central ζ=α guarantee is not valid for general system pairs.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central guarantee is Corollary 2, ζ=β=α, which requires Assumption 5 (V+∩∂Λ nonempty) and Assumption 4. In the constructive method of Section V-A, C is assembled from the IP and semipermeable surfaces L integrated retrograde from the BIP via (13)-(16); the asserted safety property holds only if these surfaces meet in a nonleaking closed barrier K (Section V-A.3, [18, Sec. 4.3]). No theorem in the manuscript proves that such a K exists for arbitrary f, U_l, U_h, and α. Assumption 6 merely postulates that the relevant ξ_L contains a part ξ_K; Remark 4 asserts ν(β) exists for a 'proper pair' without proof; and the optimization (18)-(22) is silent when no barrier exists. If the true maximal invariant set lies strictly in the interior of Λ, then the computed WTE is strictly less than α, so O.1 is infeasible for that margin, and the minimal-intervention controller of Theorem 3 cannot deliver the claimed ζ=α. Section VI.E explicitly states that the coincidence with FaSTrack TEBs is only a 'hypothesis' with a formal proof 'still under development,' which concedes that the existence claim is not established. This is a structural existence premise, not a numerical-accuracy issue.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":15509,"tokens_out":4190,"duration_ms":46831,"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":[{"comment":"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":"Section V-A.3 and Assumption 6"},{"comment":"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":"Section V-B, equations (18)-(22)"},{"comment":"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.","section":"Section VI.E"}],"minor_comments":[{"comment":"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":"Section VI-A, equations (24)-(27)"},{"comment":"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.","section":"Section VI-B, after equation (33)"},{"comment":"Reference [19] contains a typo: 'Wrigth' should be 'Wright.'","section":"References"},{"comment":"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.","section":"Section III-B and Definition 12"}],"recommendation":"major_revision","confidential_remarks":"The abstract and the conclusion overstate the guarantees relative to the manuscript's own Section VI.E, which states that a formal proof of the general equivalence is still under development. I would be comfortable with publication after the existence question is either resolved under stated assumptions or explicitly removed from the scope of the main claims, with a dedicated limitations paragraph added before acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","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.","headline":"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.","tokens_in":775,"tokens_out":920,"would_cite":true,"duration_ms":28007,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["91A23","49N70"],"pacs":[],"model":"deepseek-v4-flash","headline":"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.","keywords":["captivity-escape game","differential game of kind","safety margin","tracking error bound","worst-case tracking error","motion generation","homicidal chauffeur","safety controller"],"falsifier":"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.","tokens_in":14942,"feed_emoji":"🎯","tokens_out":9376,"duration_ms":104421,"temperature":0.7,"pith_summary":"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.","feed_headline":"Captivity-escape game lets planners hit a safety margin exactly","feed_subtitle":"Reversing pursuit-evasion yields tracking error equal to the safety margin, at far lower cost.","key_machinery":"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.","core_discovery":"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.","pith_inferences":["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."],"forward_implications":["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."],"supporting_citations":[{"why":"introduces the FaSTrack method that the paper uses as the state-of-the-art baseline for safety margin computation.","marker":"[1]"},{"why":"extends and formalizes FaSTrack and supplies the pursuit-evasion game setup that the captivity-escape game inverts.","marker":"[2]"},{"why":"gives the SOS-optimization alternative for addressing the safety margin problem that the paper compares for conservatism.","marker":"[3]"},{"why":"provides the nonanticipative strategy and Hamilton-Jacobi reachability background for the game formulation.","marker":"[7]"},{"why":"surveys pursuit-evasion games of kind and semipermeable surfaces, the construction tools used for the barrier.","marker":"[14]"},{"why":"provides Isaacs's method for solving games of kind via the inward-facing part, boundary of the inward-facing part, and closed barrier.","marker":"[16]"},{"why":"supplies the analytical barrier description for the homicidal-chauffeur relative dynamics used in the numerical example.","marker":"[17]"}],"fun_headline_variants":["Captivity-escape game pins worst-case tracking error to safety margin","Game flips pursuit-evasion to meet safety margins precisely","Zero-sum game cuts planning conservatism, speeds up safety check","Homicidal-chauffeur game yields fast, exact safety margins for motion","New game-theoretic method computes safe motion plans orders faster"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["Captivity-escape game pins worst-case tracking error to safety margin","Game flips pursuit-evasion to meet safety margins precisely","Zero-sum game cuts planning conservatism, speeds up safety check","Homicidal-chauffeur game yields fast, exact safety margins for motion","New game-theoretic method computes safe motion plans orders faster"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000297,"raw_usage":{"total_tokens":1658,"prompt_tokens":820,"completion_tokens":838,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":436,"completion_tokens_details":{"reasoning_tokens":750}},"tokens_in":436,"tokens_out":838,"duration_ms":8301,"temperature":1.0,"reasoning_tokens":750,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T11:45:00.631290+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[{"cited_title":"Fastrack: A modular framework for fast and guaranteed safe motion planning,","cited_arxiv_id":null,"evidence_quote":"introduces the FaSTrack method that the paper uses as the state-of-the-art baseline for safety margin computation."},{"cited_title":"Fastrack: A modular framework for real-time motion planning and guaranteed safe tracking,","cited_arxiv_id":null,"evidence_quote":"extends and formalizes FaSTrack and supplies the pursuit-evasion game setup that the captivity-escape game inverts."},{"cited_title":"Robust tracking with model mismatch for fast and safe planning: An sos optimization approach,","cited_arxiv_id":null,"evidence_quote":"gives the SOS-optimization alternative for addressing the safety margin problem that the paper compares for conservatism."},{"cited_title":"A time-dependent hamilton-jacobi formulation of reachable sets for continuous dynamic games,","cited_arxiv_id":null,"evidence_quote":"provides the nonanticipative strategy and Hamilton-Jacobi reachability background for the game formulation."},{"cited_title":"Pursuit-Evasion Games,","cited_arxiv_id":null,"evidence_quote":"surveys pursuit-evasion games of kind and semipermeable surfaces, the construction tools used for the barrier."},{"cited_title":"Isaacs,Differential Games, New ed Edition","cited_arxiv_id":null,"evidence_quote":"provides Isaacs's method for solving games of kind via the inward-facing part, boundary of the inward-facing part, and closed barrier."},{"cited_title":"The Game of Two Identical Cars: An Analytical Description of the Barrier,","cited_arxiv_id":null,"evidence_quote":"supplies the analytical barrier description for the homicidal-chauffeur relative dynamics used in the numerical example."}],"review_version":1}