{"id":"9b5d538b-b68a-4e95-b5dc-a3ffd6d22fa1","arxiv_id":"2607.10986","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Minimax receding-horizon MPC plus a discrete CBF safety filter lets a pursuer swarm switch capture, perimeter defense, and neutralization modes by retuning shared costs and barrier weights alone.","lead":"A hierarchical controller lets robot swarms switch between capturing, shielding, or neutralizing an adversary by retuning shared game costs and safety margins, without redesigning the control stack. It matters for counter-UAV and perimeter defense, where rules of engagement change mid-mission.","discovery_kind":"extension","skeptic_critique":{"model":"grok-4.5","headline":"Offline weight/CBF pairs do not yet demonstrate live fluid switching under the same stack.","rationale":"The reader's weakest-assumption diagnosis is exactly the soft spot: the central adaptability claim is only partially evidenced by static offline weight/CBF settings, while live switching is deferred. Methods (minimax MPC + discrete CBF-QP) and multi-regime sims are coherent and code is released, so the architecture demonstration remains accept-shaped once live switches, baselines, and weight tables are added. No stronger internal inconsistency appears; the concern is evidentiary scope, not soundness of the offline results. Verdict therefore stays CONDITIONAL.","tokens_in":12279,"tokens_out":442,"duration_ms":5451,"concrete_test":"Using the released code, run a single continuous 2D double-integrator episode that mid-horizon switches the active stage cost (J_PE ↔ J_AD) and the CBF pair (D_safe,pe, w_slack) at a fixed time, without re-initializing agents or solvers. Report capture/breach rates, inter-agent CBF violations, and QP feasibility over ≥20 Monte-Carlo trials; if any metric degrades >15% relative to the corresponding static-mode baseline, the 'fluid switch without structural change' claim is unsupported.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The abstract and §I claim that 'by simply altering the weights of the shared zero-sum payoff and CBF constraints, the swarm can fluidly switch' among capture, perimeter defense, and neutralization 'without structural changes to the control logic.' §III-A and §IV-B implement ROE by coupling two distinct MPC stage costs (J_PE vs J_AD) with static offline choices of D_safe,pe and w_slack. The paper itself states that experiments use only static offline configurations and that online reactive switching is future work (§IV-B). Thus the load-bearing premise—that the same hierarchical stack supports fluid live mode changes by weight/CBF retuning alone—is asserted but not shown; the demonstrated result is that different static parameterizations induce different behaviors.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper proposes a hierarchical multi-agent control stack for adversarial pursuit–evasion that separates strategic planning from safety. Planning is cast as a finite-horizon zero-sum game solved by iterative minimax (alternating best-response) MPC with two stage-cost designs: a pursuit–evasion cost with relative-distance and transverse-velocity (cutoff) penalties, and an area-denial/perimeter-defense cost that tracks boundary points and denies evader progress. Execution is a discrete-time velocity-aware CBF-QP filter that enforces inter-pursuer collision avoidance and pursuer–evader standoff with slacks. Rules of engagement (capture, shield/deflect, neutralize) are said to be realized by retuning shared payoff weights and CBF margins/slacks without redesigning the control logic. Simulated 2D double-integrator Monte Carlo capture grids (100 games), qualitative perimeter-defense and ROE demos, and a 3D quadrotor ROS/RViz capture sequence support feasibility; code is released.","tokens_in":12594,"tokens_out":976,"duration_ms":18005,"significance":"If the hierarchical stack and ROE parameterization hold up under stronger evaluation, the work is a useful systems contribution for engagement-aware multi-robot defense: it cleanly decouples convex safety filtering from nonconvex strategic planning, uses a geometrically motivated transverse-velocity penalty rather than pure reactive formations, and shows the same pipeline on both planar double integrators and linearized 6-DOF quadrotor dynamics with real-time middleware. Public source code, Monte Carlo capture heatmaps, and an Appendix B argument that the velocity-aware barrier remains a valid CBF under quadrotor thrust are concrete strengths. The main novelty claim—fluid capture/shield/neutralize switching by weight and CBF retuning alone without structural change—is of clear operational interest for counter-UAV and perimeter defense, but currently rests more on static offline parameterizations than on demonstrated live mode changes.","major_comments":[{"comment":"Abstract and §I claim that “by simply altering the weights of the shared zero-sum payoff and CBF constraints, the swarm can fluidly switch” among capture, perimeter defense, and neutralization “without structural changes to the control logic.” §III-B instead introduces two distinct stage-cost structures, J_PE (relative distance + transverse velocity cutoff) and J_AD (boundary tracking + progress denial with different terminal terms), not a single shared payoff with reweighted coefficients. §IV-B further states that engagement demos use static offline configurations and that online reactive switching is future work. The load-bearing claim should be restated to match what is shown (different static MPC objectives + CBF margins induce different modes), or live online weight/CBF switching under one fixed cost structure should be demonstrated with quantitative metrics.","section":null},{"comment":"§III-B and §IV: quantitative evaluation is uneven. Capture is supported by 100-game velocity and pursuer-count grids (Figs. 2–3), but perimeter defense (Fig. 4) and ROE standoff vs. neutralization (Fig. 6) are qualitative snapshots only—no breach rates, time-to-breach, standoff-violation rates, or inter-pursuer collision counts under matched seeds. Without those metrics (and without baselines such as pure distance-minimizing MPC, heuristic ring formations, or CBF-only controllers), the claim of robust performance “across varying rules of engagement” is not yet load-bearing.","section":null},{"comment":"§III-B: the zero-sum game is approximated by a fixed number of alternating best-response NLPs (“Nash iterations”) with frozen opponent trajectories. There is no report of iteration count, residual of the best-response map, sensitivity to initialization, or comparison to a true simultaneous minimax solve. Because the abstract markets “iterative minimax MPC” that anticipates and blocks the evader, the approximation quality of this procedure should be characterized at least empirically (e.g., capture rate vs. number of iterations, or open-loop cost gap).","section":null},{"comment":"§IV-C (3D quadrotor): dynamics are linearized once per cycle and frozen over the horizon (RTI). Capture is shown only as a single trajectory sequence (Fig. 5). Given the paper’s emphasis on scalability to nonlinear 6-DOF agents, a small Monte Carlo or parameter sweep (as in 2D) and explicit statement of linearization/horizon settings needed for reliable capture would better support the 3D claim.","section":null}],"minor_comments":[],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The one thing worth knowing is that this is a clean engineering packaging of minimax receding-horizon MPC plus a discrete velocity-aware CBF filter, with code and multi-regime sims, not a new game-theoretic result. The abstract’s “fluidly switch by simply altering weights” claim is stronger than the experiments: they change the stage cost (J_PE vs J_AD) and freeze offline D_safe / slack values; live reactive switching is explicitly future work.\n\nWhat is new is treating ROE as joint design objects on the shared zero-sum cost and the CBF margins/slacks, so the same two-layer stack can be retuned from capture to perimeter defense to neutralization. They do this carefully: transverse velocity cutoff in the PE cost, polygonal boundary tracking for area denial, alternating best-response NLPs, and a braking-distance barrier that keeps the outer MPC convex. Appendix B checks that thrust appears in ḣ for the quadrotor linearization. Monte Carlo capture grids (100 games, two initializations, pursuer-count sweeps) and a ROS 2 / RViz 3D demo are reproducible enough to be useful. Self-citations are background tools, not the measured outcome.\n\nSoft spots are real but proportional. No quantitative baselines against other PE or CBF-MPC controllers. Perfect-state deterministic dynamics only. Free parameters (weights, D_safe, γ, Nash iterations, RTI linearization) are numerous and not tabulated for the mode figures. The stress-test note is right: offline static configurations do not yet demonstrate live fluid switching under the same stack. That does not break the architecture; it just means the strongest sentence in the abstract is aspirational.\n\nThis is for multi-robot safety and counter-UAV control people who want a working hierarchical template they can retune. It is not for pure differential-game theorists. The math is coherent, the data support feasibility, and the citation pattern is normal for the subfield. I would send it to peer review; a serious referee can demand live mode-switch demos, baselines, and weight tables without the paper collapsing. Worth engaging if you care about deployable adversarial swarm stacks.","headline":"Solid hierarchical PE stack with released code; the fluid ROE claim is oversold because only static offline weight/CBF settings are shown.","tokens_in":13226,"tokens_out":521,"would_cite":false,"duration_ms":5357,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"A single hierarchical controller lets a swarm switch from capture to perimeter defense to neutralization by retuning cost weights and safety margins, without rewriting the control law.","keywords":["pursuit-evasion","rules of engagement","minimax model predictive control","control barrier functions","multi-agent systems","perimeter defense","hierarchical control"],"falsifier":"Run the identical controller stack while a high-level supervisor changes the payoff weights and CBF standoff/slack parameters mid-flight; if the swarm fails to transition cleanly among capture, shielding, and neutralization without collisions or loss of feasibility, the central adaptability claim fails.","tokens_in":13180,"feed_emoji":"🎯","tokens_out":611,"duration_ms":5901,"temperature":0.7,"pith_summary":"This paper claims that multi-agent pursuit-evasion does not need a new controller for every mission. Capture, shielding a perimeter, and neutralizing an adversary can all be expressed as different weights on one shared zero-sum payoff plus different margins on a safety filter. The high layer runs an iterative minimax model-predictive controller that anticipates the opponent; the low layer is a discrete-time control-barrier-function quadratic program that enforces collision and standoff constraints while preserving convexity. In simulation the same stack works for 2-D double-integrator agents and for 3-D nonlinear quadrotors, and simply changing those weights and margins produces the three distinct rules of engagement. A sympathetic reader cares because the design separates strategy from safety so that rules of engagement become tunable parameters rather than hard-coded behaviors.","feed_headline":"One controller stack switches capture, shield, neutralize","feed_subtitle":"Retune cost weights and safety margins; the same MPC-CBF hierarchy changes mission without rewrite","key_machinery":"The engagement-aware hierarchy: an iterative minimax MPC that optimizes a shared zero-sum cost containing either relative-distance-plus-transverse-velocity terms (pursuit) or boundary-tracking and progress-denial terms (area denial), cascaded with a discrete-time velocity-aware CBF quadratic program that softens inter-agent and pursuer-evader distances via tunable slacks.","core_discovery":"A hierarchical architecture that treats rules of engagement as control-design objects can realize capture, perimeter defense, and neutralization with one receding-horizon minimax planner and one discrete-time CBF safety filter; only the shared zero-sum payoff weights and the CBF standoff/slack parameters need to change.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["One minimax-MPC plus CBF switches capture, shield, neutralize","Retune payoff weights: same stack flips pursuit to perimeter denial","Hierarchical MPC-CBF realizes capture, shield or neutralize on demand","Engagement rules as tunable costs: one controller for three tactics","Zero-sum MPC and discrete CBF adapt swarm from chase to area denial"],"cache_read_input_tokens":128,"weakest_assumption_plain":"The demonstrations of mode switching use only static, offline choices of weights and safety margins; the claim that the same stack will switch modes online without retuning is left as future work.","fun_headline_variants_meta":{"raw":{"variants":["One minimax-MPC plus CBF switches capture, shield, neutralize","Retune payoff weights: same stack flips pursuit to perimeter denial","Hierarchical MPC-CBF realizes capture, shield or neutralize on demand","Engagement rules as tunable costs: one controller for three tactics","Zero-sum MPC and discrete CBF adapt swarm from chase to area denial"]},"model":"grok-4.5","effort":"low","cost_usd":0.003476,"raw_usage":{"total_tokens":1122,"prompt_tokens":718,"num_sources_used":0,"completion_tokens":93,"cost_in_usd_ticks":34760000,"prompt_tokens_details":{"text_tokens":718,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":311,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":718,"tokens_out":93,"duration_ms":3106,"temperature":1.0,"reasoning_tokens":311,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-14T07:50:26.022554+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Run the identical controller stack while a high-level supervisor changes the payoff weights and CBF standoff/slack parameters mid-flight; if the swarm fails to transition cleanly among capture, shielding, and neutralization without collisions or loss of feasibility, the central adaptability claim fails.","supporting_citations":[],"review_version":1}