{"id":"6abfbde5-75a4-4a58-9a8e-cc7c78292386","arxiv_id":"2607.21043","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"A GNEP-and-PSO planner for two-vehicle intersection interaction ran under 50 ms and produced cooperative real-world left-turn trajectories, yet the reported 'Nash cost' is just the minimized sampled objective.","lead":"This paper describes a game-theoretic driving planner in which an autonomous car and a human-driven car jointly solve an intersection crossing as a two-player equilibrium problem; a particle-swarm optimizer finds a candidate solution in under 50 milliseconds. The authors tested it in a real Renault Zoé and report cooperative left-turn trajectories, but the equilibrium verification is a sampled lower-bound objective and no baseline comparisons are provided.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported 'Nash Cost 0.0000' (Table I) is the PSO's own minimized objective Ĵ — a random-sample lower bound on the true deviation J (Eqs. 18–22) — so zero Ĵ does not certify a Nash equilibrium of the continuous GNEP; the central convergence claim is unsupported.","rationale":"Load-bearing step. For the paper's central claim to hold, the optimized quantity must be a sound measure of distance to a generalized Nash equilibrium. The exact J is sound in principle: with true suprema over the coupled feasible sets, J(sv,so)=0 iff (sv,so) is a global GNE. The argument breaks at the move from exact J to sampled Ĵ: Eqs. (18)–(20) and (22) replace suprema by maxima over random finite subsets, and the authors acknowledge Ĵ ∈ [0, J]. Minimizing a lower bound and reporting its value as 'Nash Cost' cannot certify equilibrium; the gap J − Ĵ is unmeasured and unmentioned. Every downstream claim — real-time Nash convergence (Table I), Fig. 4's 'convergence to Nash equilibrium', the abstract's and conclusion's equilibrium claims — inherits this gap.\n\nSecondary but reinforcing: the sampled deviation also drops the GNEP coupling (alternatives sampled from Sv/So rather than Sv(so)/So(sv)), the PSO hyperparameters and Q_i/h definitions are absent (irreproducible), the timing hardware is unspecified, and the experimental evidence is two anecdotes with no baseline. These are consistent with the reader's high correctness risk, but the load-bearing single point is the certification metric.\n\nAgreement. The reader's rationale identifies this exact issue ('the reported Nash cost is the sampled lower-bound objective being minimized by the solver itself'); the reader's stated weakest_assumption (opponent rationality/known Q_o) is a related but distinct epistemic threat to the real-world validity of the equilibrium rather than to the solver's convergence claim. Hence partial agreement.\n\nVerdict. The reader's REJECT stands. The concern is decisive: the reported evidence for the central claim measures the optimizer's own progress on a biased proxy, not the property being claimed. If the proposed dense re-evaluation showed the residual is genuinely small, the paper would merit a conditional accept with the missing method details; as written, the central claim is unsupported.","tokens_in":8522,"tokens_out":9429,"duration_ms":103205,"concrete_test":"Recompute the deviation at the PSO-reported global-best profile for the 160×160 configuration using an independent, much denser evaluation: (i) draw 10^4–10^5 Monte Carlo alternatives from Sv(so) and So(sv) (rejection-sampled so h ≤ 0 holds) and recompute Δv, Δo; and (ii) run a local continuous minimization of Qv(·,so) and Qo(sv,·) over the coupled feasible sets to seek an improving deviation. If either finds an improvement exceeding a small tolerance (e.g., 10^-2 of the cost scale), then the Table I zero is a sparse-sampling artifact and the equilibrium claim fails. A minimal screening version: report a bootstrap/confidence bound on J at the reported solution; the paper currently reports none.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that the PSO solver 'consistently converges to a valid Nash equilibrium' (Table I note) is certified only by the very quantity the solver minimizes. The exact joint criterion J(sv,so) = Δv + Δo (Eqs. 9–14) is defined by suprema over the continuous, coupled feasible sets Sv(so), So(sv) (Eq. 11). In Section III.D and Eq. (22), each supremum is replaced by a maximum over random finite samples Ŝv ⊂ Sv, Ŝo ⊂ So; the paper itself states Ĵ ∈ [0, J]. So Ĵ is a (biased) lower bound on the true deviation, and no bound on the gap J − Ĵ is given. Reporting Ĵ = 0.0000 after minimizing Ĵ is circular: with only 50–160 random samples per agent in a continuous, likely high-dimensional strategy space, a profile far from any Nash equilibrium can dominate every sampled alternative and still yield zero. The measured 'equilibrium distance' is therefore not certification. A second inconsistency compounds this: the discrete deviation (Eq. 18) draws alternatives from Sv and So, not from the coupled feasible sets Sv(so) = {sv ∈ Sv : h(sv,so) ≤ 0} defined in Eq. (5); infeasible 'improvements' may be counted, so the sampled criterion is not even a faithful discrete analog of Eq. (11). The real-vehicle section (one driver, two scenarios, no baselines, no repeats) cannot carry the certification burden.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a two-player Generalized Nash Equilibrium Problem (GNEP) formulation for autonomous driving at unsignalized intersections, where shared safety and geometric constraints couple the players' feasible sets. A joint deviation criterion J = Δv + Δo is introduced, and because the exact suprema are expensive, the authors approximate them by maxima over random finite samples, yielding Ĵ, and minimize Ĵ with a Particle Swarm Optimization (PSO) solver. The approach is evaluated in simulation benchmarks with 50–160 sampled strategies per agent and with a real autonomous Renault Zoé interacting with a human driver in two left-turn scenarios. The paper claims real-time (<50 ms) convergence to a valid Nash equilibrium and human-like, comfortable interaction-aware behavior.","tokens_in":9037,"tokens_out":6498,"duration_ms":75239,"significance":"If the central claims held, the paper would offer a practical, real-time GNEP solver for interaction-aware driving and a rare real-vehicle demonstration. The GNEP modeling with coupled feasible sets is appropriate, and the authors are transparent about Ĵ being a lower bound on the true deviation J. The real-vehicle experiments are a useful proof-of-concept. However, the reported equilibrium certification is not valid: the zero Nash cost is the minimized lower-bound estimator, not a verified deviation from equilibrium. The discrepancy between the GNEP deviation (Eq. 11) and the sampled criterion (Eq. 22) further undermines the solver's interpretation. The significance of the contribution is therefore not established in the current form.","major_comments":[{"comment":"The 'Nash Cost 0.0000' reported in Table I is Ĵ, a random-sample lower bound on the true deviation J, as the paper itself states after Eq. (22) (Ĵ ∈ [0,J]). Since PSO minimizes Ĵ, a zero value only means that no sampled alternative improves on the incumbent; it does not certify that no continuous feasible alternative improves. No gap bound, Lipschitz constant, grid-resolution argument, or independent verification is provided. With 50–160 samples in a continuous, likely high-dimensional strategy space, zero Ĵ is expected even far from equilibrium. Thus the Table I note 'The PSO solver consistently converges to a valid Nash equilibrium' is unsupported. Please provide an upper-bound verification or a formal approximation guarantee.","section":"Section III.D, Eqs. (18)–(22); Table I"},{"comment":"For a GNEP, the deviation measure must use the coupled feasible sets S_v(s_o) and S_o(s_v) defined in Eqs. (5)–(6). However, Eq. (17) and the discrete versions in Eqs. (18)–(19) and (22) maximize over the independent sets S_v and S_o. This counts infeasible unilateral deviations, so J and Ĵ are not the GNEP deviation measures. A profile with zero Ĵ may violate a coupled constraint or may fail to be a generalized Nash equilibrium. The formulation needs to enforce h(s̄_v,s_o)≤0 and h(s_v,s̄_o)≤0 in the deviation alternatives, or provide an argument that the independent sets coincide.","section":"Section III.C–D, Eq. (11) versus Eqs. (17)–(22)"},{"comment":"The strategy spaces and sampling procedure are underspecified. The dimensions n,m of z_v,z_o are never given; the '50×50' grids are described only as 'strategies per agent' without specifying how they are sampled, over what horizon, or under what constraints. PSO hyperparameters (ω,c1,c2,P,number of iterations) are not stated, and the 'Evaluations Used' column is undefined. The runtime and convergence results are therefore not reproducible, and the claimed <50 ms convergence cannot be independently assessed. Please provide the full algorithm configuration and search-space details.","section":"Section IV.A and Table I"},{"comment":"The real-vehicle evaluation consists of one human driver, two scenario descriptions, and qualitative trajectory plots. There are no repeated trials, no baseline comparisons, and no quantitative measures of comfort, safety margin, or success rate. This evidence cannot support the general claims of 'comfortable, human-like trajectories' or robustness to behavioral variability. The section should be framed as a demonstration rather than a validation, and ideally supplemented with statistical measures and baselines.","section":"Section V"}],"minor_comments":[{"comment":"The minimization variables are written as “ās_v,ās_o,” but the objective J depends on s_v,s_o; as written the problem is ill-posed. This should be corrected to minimize over (s_v,s_o) ∈ S_v × S_o, and the notation throughout should be harmonized.","section":"Eq. (16)"},{"comment":"The phrase 'aggregation f two deviation measures' is missing the word 'of'.","section":"Section III.D, sentence after Eq. (12)"},{"comment":"Only the convergence profile for Use Case 2 is shown; include the corresponding profile for Use Case 1 for comparability.","section":"Figure 4"},{"comment":"Define the 'Evaluations Used' column and state whether the entries are function evaluations per PSO run or a fraction of a precomputed grid. Also clarify whether the same random sample sets are reused across particles and runs.","section":"Table I"},{"comment":"The notation for the approximate criterion is inconsistent: ĶJ appears in Eq. (20), bJ in Eq. (22), and 'Nash Cost' in Table I. Please unify.","section":"Throughout"}],"recommendation":"reject","confidential_remarks":"The core problem is not presentation but a fundamental validation gap: the equilibrium certificate is the very lower-bound objective minimized by the solver. Even if an upper-bound verification were added, the independent-strategy-set inconsistency in Eq. (17) relative to Eq. (11) would require a major reformulation of the criterion. The real-vehicle experiments are far too limited to compensate. In my view the central claim is not established in this manuscript, and a substantially new evaluation would be needed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Punchline: This is a decent engineering effort with a real vehicle demo, but the claim that the PSO solver converges to a Nash equilibrium is not supported by the evidence as presented.\n\nWhat's new and what works: The specific combination of a generalized Nash equilibrium formulation with shared safety constraints, a PSO solver that minimizes a sampled joint-deviation criterion, and a test-track demonstration with a real autonomous vehicle and a human driver is not in the existing literature. The authors honestly state that the sampled deviation is a lower bound on the true criterion, and the runtime figures are plausible. The two driving scenarios, though limited, show reasonable, non-aggressive behavior, which is a useful engineering existence proof.\n\nSoft spots: The key problem is the equilibrium certificate. Table I reports Nash cost 0.0000, but that is the value of Ĵ, the sampled lower bound that PSO is minimizing. A zero lower bound does not imply the profile is a true generalized Nash equilibrium; the gap could be arbitrarily large. No bound on the gap is given, so the claim that the solver 'consistently converges to a valid Nash equilibrium' is not supported. In addition, Eq. (18) samples alternatives from Sv and So, not from the coupled feasible sets Sv(so) and So(sv) used in the original deviation measures. That means infeasible improvements may be counted, so the sampled criterion is not even a faithful discrete version of the true deviation. This is a load-bearing flaw, not a cosmetic one. The real-vehicle experiments are also thin: one human driver, two scenarios, no baselines, no repeated trials, and no uncertainty quantification. Finally, the paper omits nearly all parameter values (cost weights, PSO coefficients, sampling strategy), so the work cannot be independently reproduced.\n\nWho it's for: This paper will be of interest to researchers working on game-theoretic planning for automated vehicles and on metaheuristic equilibrium solvers. It is a plausible engineering application, not a theoretical contribution.\n\nRecommendation: I would send this to peer review, but a referee should require a proper continuous-space equilibrium residual (or comparison to a known solver) and a full description of the method. As written, the central convergence claim is overstated. With those fixes, it could be a solid applied paper.","headline":"Real-vehicle GNEP/PSO planner that runs under 50 ms, but the equilibrium certificate is the very objective the solver minimizes, so the central claim does not hold as written.","tokens_in":9443,"tokens_out":4987,"would_cite":false,"duration_ms":52989,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A real autonomous vehicle can negotiate with a human driver by solving a shared-constraint Nash equilibrium in real time, demonstrated at a four-way intersection.","keywords":["generalized Nash equilibrium","game-theoretic planning","autonomous driving","mixed traffic","particle swarm optimization","intersection negotiation","interaction-aware planning","real-time motion planning"],"falsifier":"Record the human driver's actual trajectory during the experiment and compute, retrospectively, whether the human could have lowered their own measured cost by deviating from the strategy the planner attributed to them; if a deviation with negative cost exists at the Nash point, the equilibrium claim is falsified. A cleaner test would run the same planner against a driver with a deliberately different objective (aggressive, gambling, or altruistic) and check whether J remains near zero; if it does not, the model's assumption of shared cost structure fails.","tokens_in":8452,"feed_emoji":"🚗","tokens_out":5759,"duration_ms":62067,"temperature":0.7,"pith_summary":"This paper aims to show that interaction-aware driving can be cast as a two-player Generalized Nash Equilibrium Problem (GNEP), where each vehicle's feasible strategies are coupled by shared safety and geometry constraints. A Particle Swarm Optimization (PSO) solver minimizes a joint deviation criterion—the sum of each driver's maximum possible cost improvement from unilaterally changing strategy—and finds an equilibrium in under 50 ms for strategy grids up to 160×160. The authors validate the framework on a closed test track with an autonomous vehicle and a human-driven vehicle performing simultaneous left turns, reporting comfortable, human-like trajectories. If correct, this demonstrates that game-theoretic coordination can run online in real traffic rather than remaining a simulation-only method.","feed_headline":"Game-theory planner runs on a real car: Nash equilibrium in 50 ms","feed_subtitle":"A particle-swarm solver reaches equilibrium in under 50 ms, and the real car's left turns look human-like.","key_machinery":"The central object is the Generalized Nash Equilibrium Problem with shared inequality constraints h(s_v, s_o) ≤ 0, which makes each player's feasible strategy set depend on the opponent's choice. The equilibrium is characterized by the joint deviation criterion J(s_v, s_o) = Δ_v + Δ_o, where each Δ is the supremum (approximated by a maximum over sampled strategies) of the cost improvement a player could gain by unilaterally deviating. A Particle Swarm Optimization swarm treats each particle as a candidate strategy pair and minimizes the sampled version Ĵ, with personal and global best defined as the profiles with the smallest estimated deviation from equilibrium. This construction converts t","core_discovery":"The central claim is that a Generalized Nash Equilibrium, in which players' feasible sets depend on each other's strategies, is the right model for mixed-traffic interaction, and that a stochastic sampling version of the deviation-based criterion J = Δ_v + Δ_o can be minimized by PSO to yield an equilibrium online. Using random samples of alternative strategies, the solver estimates the largest unilateral improvement available to each player; when J drops to zero, neither driver can improve by acting alone. Benchmarks over 100 runs give mean convergence times of 19.5–48.5 ms as the strategy grid grows from 50×50 to 160×160, with a Nash cost of 0.0000 in every configuration. Real-vehicle expe","pith_inferences":["A natural extension is to estimate the opponent's cost weights online (e.g., via inverse optimal control) rather than fixing the shared structure, which would let the equilibrium track changes in a human's driving style.","The sampled deviation criterion J could be reused as a runtime safety monitor: if the estimated incentive to deviate grows during execution, the autonomous vehicle could detect model mismatch and fall back to a conservative policy.","The stochastic sampling view suggests a Monte Carlo interpretation of the equilibrium candidate; reporting the variance of J across sampled strategy sets could give a confidence measure for the claimed 0.0000 Nash cost, which currently reflects only the finite sample.","The same GNEP formulation should transfer to other mixed-traffic dyads, such as merges and pedestrian crossings, as long as the opponent's cost has the same geometric components; the paper only demonstrates intersections."],"forward_implications":["If the framework holds, an autonomous vehicle can negotiate shared conflict zones such as unsignalized intersections without predefined priority rules, by solving for a joint equilibrium online.","The PSO solver's sub-50 ms convergence for 160×160 strategy grids means the approach is compatible with typical vehicle control cycles, enabling reactive, interaction-aware replanning.","Because the joint deviation criterion reaches exactly zero in the benchmarks, the solver provides a quantitative certificate (under the sampled set) that no sampled unilateral deviation improves either player's cost.","The real-vehicle demonstrations indicate that equilibrium solutions can produce cooperative low-speed coordination and early crossing maneuvers that a human observer would read as reasonable, rather than conservative or aggressive.","Extending the two-player formulation to N > 2 players would multiply the search dimension and require encoding all pairwise collision constraints, as the authors note; the paper does not yet handle multi-vehicle intersections."],"fun_headline_variants":["Nash equilibrium on a real car: 50 ms per decision","Mixed-traffic driving solved as a game, in real time","Real car, real humans, real-time Nash equilibrium","GNEP planner: 50 ms to human-like driving","Particle swarm finds Nash equilibrium in 50 ms on road"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The approach assumes the human driver minimizes a known cost function with the same structure as the ego vehicle's; if the real driver's objectives, beliefs, or prediction horizon differ, the computed equilibrium no longer describes the interaction, and only one human driver and two scenarios were tested.","fun_headline_variants_meta":{"raw":{"variants":["Nash equilibrium on a real car: 50 ms per decision","Mixed-traffic driving solved as a game, in real time","Real car, real humans, real-time Nash equilibrium","GNEP planner: 50 ms to human-like driving","Particle swarm finds Nash equilibrium in 50 ms on road"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000923,"raw_usage":{"total_tokens":3773,"prompt_tokens":700,"completion_tokens":3073,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":444,"completion_tokens_details":{"reasoning_tokens":2989}},"tokens_in":444,"tokens_out":3073,"duration_ms":19979,"temperature":1.0,"reasoning_tokens":2989,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T08:37:55.653505+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Record the human driver's actual trajectory during the experiment and compute, retrospectively, whether the human could have lowered their own measured cost by deviating from the strategy the planner attributed to them; if a deviation with negative cost exists at the Nash point, the equilibrium claim is falsified. A cleaner test would run the same planner against a driver with a deliberately different objective (aggressive, gambling, or altruistic) and check whether J remains near zero; if it does not, the model's assumption of shared cost structure fails.","supporting_citations":[],"review_version":1}