{"id":"f3ca9869-4641-41d9-9cd8-d12e12c2ef7b","arxiv_id":"2608.06648","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A real-time Plan-and-Avoid framework resolves predicted well-clear violations by generating minimally intrusive unilateral advisories to surrounding cooperative traffic while preserving a declared priority trajectory.","lead":"This paper introduces a Plan-and-Avoid framework that plans an emergency landing trajectory and, when other traffic would violate separation, issues one of five advisory maneuvers to nearby cooperative aircraft. It is a candidate coordination layer for future uncrewed and advanced air mobility traffic management, tested on more than 900 simulated forced landings with real ADS-B traffic from Washington, D.C.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"All-575 feasibility is verified only on 1 Hz samples; continuous-time well-clear separation is unproven without denser re-check and broader 6-DOF validation.","rationale":"The reader's weakest assumption concerns cooperative intruder intent execution. That is a real limitation and is acknowledged in Section VII, but it is outside the paper's cooperative scope. The more load-bearing issue for the internal validity of the universal claim is the gap between continuous-time constraints and the 1 Hz sampled implementation. The paper explicitly discloses discrete sampling but does not state that the constraints in Eq. (44) and Eq. (51) are verified continuously or at a rate that bounds between-sample separation extrema. Since the benchmark set is large and the dynamic validation is limited to two cases, 'feasible for all 575 conflicts' should be read as feasibility on the sampled grid. This does not invalidate the architecture or the mathematical propositions; the uncertainty-deflation results in Section III.A and the robust time-window construction in Eq. (43) are sound. It does mean the headline safety claim needs a concrete dense-sampling and broader 6-DOF check before it can be taken as a continuous-time guarantee. The abstract's 'well-clear separation for all traffic' remains an overstatement because ψ2 only caps interaction-risk increase, not well-clear separation among nominal intruders; this supports the reader's conditional verdict. The verdict should remain CONDITIONAL, with the added condition that the 575-case feasibility be re-verified on a denser time grid and a larger dynamic simulation sample.","tokens_in":26386,"tokens_out":11589,"duration_ms":110187,"concrete_test":"Re-run the advisory feasibility check for all 575 conflicts using trajectories resampled at 10 Hz (or with exact interpolation of the piecewise-linear/circular segments), recomputing ψ1 (Eq. 44) and ψ2 (Eq. 49) on the denser grid. In parallel, run 6-DOF closed-loop simulations for a random sample of at least 50 of the 575 returned advisories and report the minimum separation margin distribution and any ψ1 violations. If any advisory that was feasible at 1 Hz violates ψ1≤0 under the denser grid or in 6-DOF simulation, the universal feasibility claim fails; if all remain feasible, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that PAA generates feasible advisories for all 575 conflicts and maintains well-clear separation depends on the constraints in Eq. (51) being satisfied in continuous time. The paper states in Section V that \"in practice, the trajectories are discrete, sampled at a finite frequency\" and in Section VI that \"all trajectories are sampled at a constant rate of 1 Hz.\" The robust well-clear constraint ψ1 (Eq. 44) is a max over t∈(0,T] and ξ∈W_e(t), and the QP in Eq. (51) enforces ψ1≤0 and ψ2≤0. Nothing in the implementation description indicates that these constraints are checked with interpolation or at a rate above the 1 Hz trajectory sampling. At the benchmark airspeeds (70–350 kts), an aircraft moves 36–180 m per sample, so the horizontal separation can cross the well-clear threshold between samples while remaining satisfied at sample instants. Thus the universal feasibility result is currently a property of the sampled time grid, not of continuous trajectories. The two 6-DOF closed-loop simulations (Figs. 12 and 14) are explicitly representative cases and do not cover the 575-encounter benchmark. Separately, the abstract's \"well-clear separation for all traffic\" is stronger than the evaluated interaction constraint ψ2, which only prevents increased interaction risk between the modified intruder and remaining nominal traffic. The load-bearing gap is that the headline claim rests on an unverified discretization assumption.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Plan-and-Avoid (PAA), a two-stage framework for coordinating cooperative airspace traffic around a declared priority trajectory. The Plan stage extends a previously developed contingency landing planner with multi-agent intent and uncertainty-inflated well-clear constraints, while the Avoid stage generates rank-ordered unilateral advisories (halt, speed, altitude, extend, divert) for surrounding traffic, formulated as quadratic programs with a robust time-window well-clear constraint and an interaction-risk constraint. The framework is evaluated on more than 900 forced-landing scenarios built from real ADS-B data, resolving all 575 predicted conflict encounters within a reported worst-case end-to-end time of 5.7 s, with 93.5% of advisories satisfying the RTCA DO-365 35 s temporal threshold. Two representative 6-DOF dynamic simulations are also presented to check that the kinematic advisories remain dynamically realizable.","tokens_in":26552,"tokens_out":5449,"duration_ms":52966,"significance":"If the claims are fully validated, PAA is a useful and original contribution to cooperative multi-agent contingency coordination: it couples conflict-aware priority planning with an advisory hierarchy for surrounding traffic, and it is evaluated on a large public-data benchmark rather than only on hand-picked cases. The paper is honest about the suboptimality of the planner, explicitly reports cases where the Dubins baseline performs better, and makes the software openly available. The probabilistic separation-buffer propositions (Propositions 1 and 2) are mathematically sound and provide a distribution-free way to convert position uncertainty into deterministic separation margins. However, the headline claims currently outrun what is verified: the all-575 feasibility result is checked only on 1 Hz sampled trajectories, the abstract's \"well-clear separation for all traffic\" is stronger than the interaction-risk constraint actually enforced, and the reported \"worst-case\" end-to-end time mixes a mean-plus-3-sigma planning time with a worst-case advisory time. These issues are local and fixable, but they are load-bearing for the paper's central claims.","major_comments":[{"comment":"The claim that all 575 conflicts are feasibly resolved is verified only on trajectories sampled at 1 Hz, as stated in Section VI and Section V, whereas the constraint ψ1 in Eq. (44) is defined as a maximum over continuous time t and the continuous timing window ξ∈W_e(t). At the benchmark airspeeds of 70–350 kts, an aircraft moves 36–180 m between consecutive 1 Hz samples, so a well-clear violation can occur between samples even if every sample satisfies the constraint. Since the two 6-DOF simulations are explicitly representative cases, the universal \"feasible for all 575\" statement is not currently established for continuous-time trajectories. Please add a post-hoc dense re-check or an analytic inter-sample bound, and either report the continuous-time result or restrict the claim to the sampled trajectory representation.","section":"Section VI, Eq. (51)"},{"comment":"The abstract states that the framework maintains \"well-clear separation for all traffic,\" but the interaction constraint ψ2 in Eq. (49) only requires that the advised intruder trajectory not increase the cumulative and peak interaction-risk functionals I_J and J_max relative to the nominal trajectory. It does not require well-clear separation between the modified intruder and the remaining nominal traffic, and Table II only counts conflicts with the ego priority trajectory. The benchmark therefore does not evaluate well-clear separation among all pairs of traffic. Please either add a post-advisory well-clear check against all remaining traffic, or soften the claim to \"resolves predicted well-clear conflicts with the priority trajectory while not increasing interaction risk with other traffic.\"","section":"Abstract and Section IV, Eq. (49)"},{"comment":"The paper calls 5.7 s a \"worst-case end-to-end response time,\" but Section VI.E reports that path-planning runtime is the mean plus three standard deviations, while resolution-advisory runtime is the worst case across test cases. A mean-plus-3σ planning time added to a worst-case advisory time and a fixed 1 s datalink delay is not a rigorous worst-case upper bound for the end-to-end pipeline. Please report the true maximum from the distribution in Fig. 17, or rename the claim to a \"high-confidence end-to-end time\" and justify the statistical aggregation.","section":"Section VI.E and Abstract"},{"comment":"The all-575 feasibility result is computed against nominal declared trajectory intent and assumes that intruders will execute the issued advisories. The paper acknowledges in Section VII that DAA is needed as a downstream safety layer for stochastic deviations, but the abstract and conclusions do not carry this qualification. Because noncompliance or intent deviation voids the separation guarantee, the central claim should be qualified as holding only for cooperative intruders that follow their declared intent and accept the advisory; otherwise the abstract's wording is stronger than the model supports.","section":"Sections III.B and VII"}],"minor_comments":[{"comment":"The holding-point tuple is written as q=(φ, λ, h, h, χ), which contains the altitude h twice; this appears to be a typo, since the surrounding text describes an admissible altitude interval with a floor and ceiling.","section":"Section IV.E, Eq. (70)"},{"comment":"In the definition of γ̃, the first fraction (−γ_thr−γ)/(−γ_thr−γ) is identically 1, so the expression as written does not encode the intended penalty for operating outside the preferred flight-path-angle range. Please re-check the normalization formula.","section":"Section IV.E, Eq. (72)"},{"comment":"The sentence saying the final advisory state \"may lie backward in time compared to initial advisory state\" is confusing and appears inconsistent with τ_e(t_rec−t_on) = s'_rec, which requires t_rec ≥ t_on for a nonnegative extension duration. Please clarify or correct the intended meaning.","section":"Section IV.D, text near Eq. (66)"},{"comment":"References [9] and [10] appear to be the same paper by the same authors with the same title and page range; one of them should be removed or replaced with a distinct source.","section":"References"},{"comment":"The title uses \"Plan-and-Avoid\" with a hyphen while the text uses \"Plan–and–Avoid\" with en dashes; please use a single consistent spelling across the paper.","section":"Title"}],"recommendation":"major_revision","confidential_remarks":"This is a solid, honestly reported systems paper, but the abstract and conclusions promise continuous-time and worst-case guarantees that the current evaluation does not support; the revision should focus on reconciling those claims with the sampling and runtime statistics rather than on adding new algorithms."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"What you should know: this paper is a real contribution, not a repackaging. The Plan-and-Avoid architecture—conflict-aware contingency planner plus rank-ordered unilateral advisories for surrounding traffic—is not in the cited literature, which mostly does reciprocal or local avoidance. The uncertainty-deflated separation results in Propositions 1 and 2 are correct: reverse triangle inequality and a union bound, distribution-free, and they give a defensible way to turn confidence-bounded position error into deterministic buffers. The benchmark is serious: over 900 forced-landing cases, more than 140 hours of simulated flight, real ADS-B data, and openly available code. The paper also reports honestly that the planner is suboptimal and sometimes worse than the Dubins baseline (19% of cases), and it shows two 6-DOF dynamic simulations that broadly match the kinematic advisories. That kind of reporting earns credit.\n\nNow the soft spots, in proportion. The biggest one is the headline feasibility claim. The paper says PAA generates feasible advisories for all 575 conflicts, and the abstract says well-clear separation is maintained for all traffic. But the well-clear constraint in Eq. (44) is a continuous-time condition, and the implementation checks it on 1 Hz samples. At the airspeeds in the benchmark, an aircraft moves tens to hundreds of meters per sample, so horizontal separation can dip below the threshold between samples and still look fine at sample instants. The two 6-DOF validations cover only two representative cases, not the 575. So the all-575 result is currently a property of the sampled grid, not of continuous trajectories. This is a real gap, though not fatal to the architecture; the fix is to re-check with interpolation or a denser grid, and to soften the abstract to say what was actually verified.\n\nThe second issue is that the abstract's \"well-clear separation for all traffic\" exceeds what the interaction constraint ψ2 guarantees. ψ2 prevents a modified intruder from increasing its cumulative or worst-case interaction risk with the remaining nominal traffic; it does not enforce well-clear separation among the non-priority aircraft. The paper's own framing in Section IV is more careful, but the abstract overstates.\n\nThe remaining concerns are minor. The cooperative-intent assumption is load-bearing but the paper acknowledges it and explicitly positions DAA as a downstream safety layer—that is honest, not a flaw. The hand-chosen parameters k_H, k_V, gamma_thr, and t_limit get no sensitivity analysis, and the planning comparison uses only a Dubins baseline. Both are worth asking about, but they don't undermine the core idea.\n\nBottom line: the central architecture holds up; the universal guarantee does not. This deserves a serious referee, not a desk reject, but it needs revision before acceptance. I'd want the discretization gap addressed and the claims calibrated to what the evaluation actually shows.","headline":"A genuine two-stage priority-trajectory coordination architecture with sound uncertainty math and a large real-data benchmark, but the all-575 well-clear claim is only established on a 1 Hz sampled grid, not in continuous time.","tokens_in":27188,"tokens_out":1807,"would_cite":true,"duration_ms":20316,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A two-stage plan-and-avoid loop resolves every predicted well-clear conflict in its 575-encounter benchmark, with a worst-case end-to-end time of 5.7 seconds.","keywords":["multi-agent airspace coordination","well-clear separation","contingency landing planning","conflict resolution advisory","uncertainty-deflated separation","ADS-B benchmark","priority trajectory","detect-and-avoid"],"falsifier":"Run the released implementation on the same 575-encounter benchmark with an added simulation layer in which intruders execute their advisories with small tracking errors, such as a 5 knot speed error or a 0.5 degree heading error, while the ego stays within its declared speed bounds; any resulting loss of well-clear separation would show the advisory pipeline is not robust to execution noise within the assumed envelope.","tokens_in":26069,"feed_emoji":"✈️","tokens_out":4583,"duration_ms":39491,"temperature":0.7,"pith_summary":"The paper claims that a declared priority trajectory, such as an emergency landing path, can be preserved in shared airspace by combining a conflict-aware planner with a fast advisory layer that modifies only the surrounding cooperative traffic. It reports that in a benchmark built from real ADS-B traffic over Washington, D.C., all 575 predicted loss-of-well-clear encounters received feasible resolution advisories, with a worst-case end-to-end time of 5.7 seconds on a personal computer and 93.5% of advisories issued at least 35 seconds before the predicted conflict. The significance is that an aircraft with limited maneuverability does not have to absorb the burden of separation; nearby aircraft can be given simple, explainable instructions within seconds. The framework is explicitly positioned as complementary to Detect-and-Avoid, which remains as a downstream safety net for unexpected deviations.","feed_headline":"All 575 predicted conflicts resolved within 5.7 seconds","feed_subtitle":"Plan-and-Avoid keeps the priority aircraft's path untouched and reroutes cooperative traffic with feasible advisories.","key_machinery":"The load-bearing device is the uncertainty-deflated separation metric $\\delta_H(t)=\\max(0,\\|\\hat{r}_H(t)\\|-\\Delta\\bar{e}_H)$ and $\\delta_V(t)=\\max(0,|\\hat{r}_V(t)|-\\Delta\\bar{e}_V)$, which converts confidence-bounded position errors into deterministic buffers using the reverse triangle inequality and the union bound. This lets both the planner and the advisory optimizer check well-clear separation with a distribution-free confidence guarantee, replacing probabilistic position uncertainty with inflated separation thresholds. The advisory side is carried by a rank-ordered action set $\\mathcal{R}=\\{R_{Halt},R_{Speed},R_{Alt},R_{Extend},R_{Divert}\\}$, where each action is a trajectory transformation solved as a quadratic program subject to two constraints: $\\psi_1$ enforces robust well-clear separation with the priority ego including its speed-tracking time window, and $\\psi_2$ requires that the advised trajectory does not increase cumulative or peak interaction risk against remaining nominal traffic.","core_discovery":"On its own terms, the paper establishes that Plan-and-Avoid (PAA) can turn an emergency landing trajectory into a system-level coordination event. The Plan stage extends a gradient-guided contingency landing planner with spatiotemporal conflict cost, using uncertainty-deflated separation buffers derived from confidence bounds on position error; the Avoid stage then issues rank-ordered unilateral advisories, halt, speed, altitude, extend, and divert, to conflicting intruders, each formulated as a quadratic program that must keep well-clear separation with the priority aircraft and not worsen interaction risk with other nominal traffic. The benchmark over more than 900 forced-landing cases using real ADS-B traffic from the Washington, D.C., airspace yielded feasible advisories for all 575 unique conflict encounters, with worst-case end-to-end response time 5.7 seconds including a 1 second two-way datalink delay, and 93.5% of advisories meeting the 35 second RTCA DO-365 alerting threshold.","pith_inferences":["The same architecture could invert the priority: instead of a distressed aircraft, the protected trajectory could be a time-critical delivery, a VIP route, or a corridor closure, and the advisory set would apply with unchanged logic.","The interaction-risk constraint $\\psi_2$ is what makes the approach scalable; without it, resolving one conflict could cascade into new conflicts, and the paper's benchmark implicitly tests this through the requirement that all 575 encounters be resolved sequentially.","Because 6.5% of advisories fall inside the 35 second DAA horizon, those cases would already be in the reactive DAA regime; a testable improvement would be to pre-plan candidate priority trajectories before declaration, shortening the effective response time.","The 1 second optimizer limit trades optimality for latency; the paper's own 15 second comparison shows a modest total benefit, suggesting that parallel or cloud computing could close the gap without violating the real-time bound."],"forward_implications":["Conflict-aware search planning reduces loss-of-well-clear exposure by 45.5% in accumulated severity and increases conflict-free cases from 50.4% to 60% compared with a Dubins baseline.","All 575 unique conflict encounters in the benchmark received feasible advisories, and speed regulation was the most common resolution (36.2%), suggesting that many conflicts can be resolved by modest timing changes.","The 1 second optimizer limit preserves the overall distribution of advisory types compared with 15 seconds of optimization, with speed remaining dominant and only a moderate shift from extend to divert advisories.","Worst-case end-to-end runtime of 5.7 seconds on a personal computer, including path planning, advisory generation, and a 1 second datalink delay, means the coordination loop closes within seconds of the priority trajectory being declared.","A large majority, 93.5%, of advisories are issued at least 35 seconds before the predicted conflict, the RTCA DO-365 alerting threshold."],"supporting_citations":[{"why":"Supplies the base gradient-guided contingency landing planner that the Plan stage extends with multi-agent conflict cost.","marker":"[36]"},{"why":"Defines well-clear and the 35 second modified tau alerting threshold used to evaluate advisory lead time.","marker":"[32]"},{"why":"Provides the real ADS-B traffic data used to create the benchmark environment.","marker":"[47]"},{"why":"Supplies the horizontal and vertical position error bounds used for the uncertainty buffers.","marker":"[48]"},{"why":"Supplies the ADS-B position error bounds used for the uncertainty buffers.","marker":"[49]"},{"why":"Provides the S-turn Dubins path solver used for extend advisories.","marker":"[45]"},{"why":"Provides the Dubins curve construction underlying the path solver.","marker":"[46]"},{"why":"Provides the NASA Generic Transport Model used in dynamic simulations to verify that advisories are dynamically realizable.","marker":"[50]"},{"why":"Provides the flight envelope and dynamics model used for the transport-class aircraft in dynamic simulations.","marker":"[51]"}],"fun_headline_variants":["Plan-and-Avoid resolves 575 conflicts in 5.7 seconds","Real-time collision avoidance for aircraft: 575 cases solved","575 aircraft conflicts handled in under 6 seconds","Priority flight path kept safe: PAA solves 575 encounters","575 conflicts, 5.7s response: PAA preserves priority route"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The framework assumes intruders share accurate time-stamped intent and will follow the issued advisories; if an intruder deviates from its declared intent or ignores an advisory, well-clear separation is not guaranteed, which the paper acknowledges by positioning Detect-and-Avoid as a downstream safety layer.","fun_headline_variants_meta":{"raw":{"variants":["Plan-and-Avoid resolves 575 conflicts in 5.7 seconds","Real-time collision avoidance for aircraft: 575 cases solved","575 aircraft conflicts handled in under 6 seconds","Priority flight path kept safe: PAA solves 575 encounters","575 conflicts, 5.7s response: PAA preserves priority route"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000221,"raw_usage":{"total_tokens":1498,"prompt_tokens":1039,"completion_tokens":459,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":655,"completion_tokens_details":{"reasoning_tokens":372}},"tokens_in":655,"tokens_out":459,"duration_ms":4500,"temperature":1.0,"reasoning_tokens":372,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T04:11:33.214537+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the released implementation on the same 575-encounter benchmark with an added simulation layer in which intruders execute their advisories with small tracking errors, such as a 5 knot speed error or a 0.5 degree heading error, while the ego stays within its declared speed bounds; any resulting loss of well-clear separation would show the advisory pipeline is not robust to execution noise within the assumed envelope.","supporting_citations":[{"cited_title":"Airspace-aware contingency landing planning,","cited_arxiv_id":null,"evidence_quote":"Supplies the base gradient-guided contingency landing planner that the Plan stage extends with multi-agent conflict cost."},{"cited_title":"Bringing up OpenSky: A large-scale ADS-B sensor network for re- search,","cited_arxiv_id":null,"evidence_quote":"Provides the real ADS-B traffic data used to create the benchmark environment."},{"cited_title":"Airworthiness Approval of Posi- tioning and Navigation Systems,","cited_arxiv_id":null,"evidence_quote":"Supplies the horizontal and vertical position error bounds used for the uncertainty buffers."},{"cited_title":"Airworthiness Approval of Automatic Dependent Surveillance– Broadcast OUT Systems,","cited_arxiv_id":null,"evidence_quote":"Supplies the ADS-B position error bounds used for the uncertainty buffers."},{"cited_title":"Emergency Flight Planning Applied to Total Loss of Thrust,","cited_arxiv_id":null,"evidence_quote":"Provides the S-turn Dubins path solver used for extend advisories."},{"cited_title":"Nasa langley’s airstar testbed: A subscale flight test capability for flight dynamics and control system experiments,","cited_arxiv_id":null,"evidence_quote":"Provides the NASA Generic Transport Model used in dynamic simulations to verify that advisories are dynamically realizable."},{"cited_title":"Design of a flight envelope protection system for nasa’s transport class model,","cited_arxiv_id":null,"evidence_quote":"Provides the flight envelope and dynamics model used for the transport-class aircraft in dynamic simulations."}],"review_version":1}