{"id":"409c1fa4-5efc-4c18-9c07-8ad080131076","arxiv_id":"2411.14567","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"V2V intention sharing lets each lane-changing vehicle solve an optimal control problem, and experiments show up to 26.8 percent fuel savings over a prediction-based baseline, with average savings of 11.2 to 15.5 percent.","lead":"Four connected cars that share their intended paths during a lane change saved fuel compared with cars that only predict each other's moves. Software and vehicle-in-the-loop experiments measured 11 to 27 percent fuel savings in some scenarios.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Algorithm 1 performs one sequential solve per agent per step against stale intentions, so it never establishes the fixed point x*_¬v that Theorem 1 requires; the open-loop plans are not shown to be a GNE.","rationale":"The reader identified the same load-bearing weakness: Theorem 1 presumes a fixed point x*_¬v that Algorithm 1 never computes, because it performs one solve per agent against potentially stale intentions. My reading confirms this is the most serious gap. The paper's novelty rests on the claim that V2V intention sharing reduces the game to an OCP whose solution is a GNE; without a fixed-point computation or convergence argument, the theoretical contribution is not established. The empirical energy savings are plausible and the testbed is a real contribution, but they do not validate the equilibrium label. A concrete test—checking the Nikaido-Isoda condition on the actual one-pass plans, or comparing against an inner fixed-point iteration—would settle whether the gap is formal or material. The MIQP nonconvexity is a secondary concern that compounds the first, but the stale-intention issue alone is decisive. Since the reader already requested conditional acceptance, my assessment does not change the verdict.","tokens_in":14611,"tokens_out":4032,"duration_ms":38825,"concrete_test":"Instrument the SiL testbed to record, at a fixed sampling instant, the shared intentions before and after each agent's solve in Algorithm 1. Then evaluate Definition 3 for the final one-pass joint plan: fix all other agents' final plans and re-solve each agent's MIQP unilaterally; if any single-agent deviation lowers that agent's cost, the joint plan is not a GNE. As a stronger check, run an inner fixed-point iteration (re-solve all agents repeatedly until shared intentions stop changing) for the same initial condition. If the one-pass plan differs from the converged fixed point, or if the fixed-point iteration fails to converge, the paper's claim that the algorithm 'arrives at a generalized Nash equilibrium' is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central theoretical claim (Lemma 1 and Theorem 1, Section II-A) is that if each agent solves an OCP with the optimal future states of others x*_¬v, the joint solution is a GNE. The proof is coherent only at a fixed point where every agent's plan is a best response to the same x*_¬v. Algorithm 1 does not compute such a fixed point: at each sampling instant it loops over agents once, solving each agent's MIQP against intentions broadcast in the previous iteration (or against predict() at initialization), then immediately updates that agent's intention. Thus agent v's plan is a best response to stale plans of others, not to the plans those agents broadcast after their own solves in the same loop. Nothing guarantees the one-pass joint solution satisfies Definition 3 (Nikaido-Isoda = 0), so the label 'open loop GNE' in Section III-C is not justified. The paper itself concedes closed-loop GNE cannot be claimed, but the problem is deeper: even the open-loop plans are not shown to be a GNE. Additionally, the MIQP contains an integer lane command u_l, so the OCP is nonconvex; Theorem 1 in [10] requires convex control sets and compact state sets, and the paper does not verify these hypotheses for the implemented model (2)-(12). The empirical energy savings may be real, but they do not establish that the mechanism claimed—arriving at a GNE via intention sharing—is what produces them.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper formulates multi-agent energy-efficient motion planning in interactive lane-change scenarios as a generalized Nash equilibrium problem (GNEP). It argues that when each agent receives the optimal intended future states of the other agents over the horizon via V2V communication, the dynamic GNEP reduces to an optimal control problem per agent, and that the resulting OCP solutions constitute a generalized Nash equilibrium. The method is implemented as a distributed mixed-integer quadratic program MPC (GNEP-MPC) and compared against a baseline MPC that predicts neighboring vehicles without V2V intentions. Experiments are conducted first in software-in-the-loop and then in vehicle-in-the-loop with a real Mazda CX-7 embedded in a mixed-reality environment. The reported results claim energy and travel-time improvements of GNEP-MPC over the baseline, including up to 26.8% fuel savings in vehicle-in-the-loop tests.","tokens_in":15057,"tokens_out":3962,"duration_ms":42608,"significance":"If the claims are accepted, the paper offers a useful experimental demonstration that V2V intention sharing can reduce energy consumption in interactive lane-change maneuvers. Its strengths are the realistic vehicle-in-the-loop setup, the comparison against a nontrivial prediction-based baseline, and the detailed description of the testbed and low-level control architecture. The theoretical contribution, however, is not established as stated: the proof that the OCP solutions form a GNE relies on a fixed-point condition that Algorithm 1 does not compute, and the theorem's convexity assumptions are not verified for the implemented MIQP. The empirical energy savings are plausible, but the paper currently presents the equilibrium mechanism as the cause without isolating it from the information advantage inherent in the comparison. The experimental reporting also lacks dispersion measures and treats one failed sub-scenario with post-hoc re-tuning. These issues are fixable within the paper's scope, but they are load-bearing for the central interpretation.","major_comments":[{"comment":"The equilibrium claim is not justified by the algorithm actually implemented. Theorem 1 proves that if every agent solves its OCP with respect to the same fixed point x*_¬v of the other agents' optimal states, then the joint solution satisfies Definition 3. Algorithm 1, however, performs a single sequential loop over agents at each sampling step and solves each agent's MIQP against intentions broadcast in the previous iteration or against predict() at initialization. Nothing in this procedure guarantees that the resulting plans are mutual best responses to the same x*_¬v, so the statement in Section III-C that 'The MPC solutions are open loop GNE' is unsupported. I recommend either computing and reporting the Nikaido-Isoda residual of Definition 3 for the actual plans, or re-describing the algorithm as a heuristic one-step intention-sharing MPC whose equilibrium properties are not established.","section":"II-A, Theorem 1 and Algorithm 1"},{"comment":"The invoked GNE theorem does not cover the implemented problem. Theorem 1 rests on hypotheses from [10] that include convex control sets and compact state sets. The implemented OCP is a mixed-integer quadratic program with an integer lane command u_l and binary collision-avoidance variables (Section II-A, Eq. (5a) and constraints (9)-(12)), so the feasible control set is nonconvex and generally noncompact. The paper does not verify that the theorem's regularity conditions are satisfied by this MIQP. At minimum, the authors should state why the integer program still satisfies the required assumptions, or explicitly restrict the theoretical claim to the convex relaxation / fixed-lane subproblem.","section":"II-A, Eqs. (5)-(12)"},{"comment":"The headline energy and travel-time improvements are reported without dispersion. Table I says the averages are over 'three trials of each sub-scenario' but does not report standard deviations, per-trial values, or the number of trials for Table II, which appears to contain a single ViL run per sub-scenario. Figure 4 shows three runs for sub-scenario 1 in SiL, but no such detail is given for the other sub-scenarios. Without measures of variability, the percent improvements cannot be assessed for statistical robustness. The authors should add error bars or per-trial tables for both SiL and ViL experiments.","section":"III-C and Table I"},{"comment":"Sub-scenario 4 in the ViL experiments reports a negative result (-12.2% energy, -3.9% travel time) and the paper states that after re-tuning the planner and tracker parameters up to 4.7% improvement could be achieved. This is a post-hoc adjustment of the system on a failure case, and as reported it is difficult to interpret. The original run and the re-tuned run should be clearly separated, with the re-tuning reported as a sensitivity/exploratory analysis rather than as part of the main comparative claim. If the re-tuning also changes the baseline or the GNEP-MPC in later sub-scenario tests, that should be disclosed.","section":"Table II and Section III-D.3"},{"comment":"The comparison does not isolate the game-theoretic mechanism. GNEP-MPC receives exact intended future states of the other agents over the horizon via V2V (Algorithm 1), while the baseline receives only the unilateral prediction model of Section III-A. The reported energy savings could therefore be due primarily to the quality and availability of the intention information, rather than to solving a GNE. A more controlled comparison would provide the baseline with the same V2V intention data while still solving a non-cooperative OCP, or would compare GNEP-MPC against an MPC that receives the same predictions but does not use the game formulation. Otherwise, the claim that 'approaching (open-loop) GNE solutions' causes the savings is not supported by the experimental design.","section":"III-C and III-D.3"}],"minor_comments":[{"comment":"The Nikaido-Isoda function is defined with a summation over agents but the notation drops the dependence of J_v on x_¬v and u_¬v. This makes the definition harder to parse; please make explicit that the minimization/deviations are over u_v only, with the other agents' controls held fixed.","section":"II-A, Definition 2"},{"comment":"The architecture figure lists the simulator, planner, and tracker components but does not identify which processes run on which machines and where the V2V communication channel is implemented. A sentence or a label in the figure clarifying the V2V data path would help reproducibility.","section":"III-B, Figure 3"},{"comment":"The text refers to an 'orange' cost curve in Figure 4, but the figure appears to be grayscale and has no legend describing which curve corresponds to which run. Please add a legend or use distinct line styles.","section":"III-C, Figure 4"},{"comment":"The fuel consumption model in Eq. (17) has a parameter p1 with units g s^2/m^2, but in the integral p1 u_t v dt the units appear inconsistent with a fuel mass rate. Please check the units or clarify the normalization.","section":"III-A, Eq. (17)"},{"comment":"The caption of Figure 10 says 'GNE approaching solutions reduce slow down behind the impeding vehicle by timely lane change.' This sentence is grammatically incomplete and should be rewritten, e.g., 'solutions approaching a GNE reduce slowing down behind the impeding vehicle ...'.","section":"III-D.3, Figure 10"},{"comment":"The paper says the four sub-scenarios are 'permutations' of (17,14,11,8) m/s, but only four of the 24 possible permutations are listed. If these four were chosen for a reason (e.g., all are 'strong' scenarios), that selection criterion should be stated explicitly.","section":"III-C, initial conditions"}],"recommendation":"major_revision","confidential_remarks":"The paper sits at the boundary of a systems journal and an experimental testbed paper. Its most defensible contribution is the vehicle-in-the-loop demonstration that V2V intention sharing can reduce fuel consumption in a lane-change scenario. The theoretical framing as a GNEP, however, is currently overclaimed relative to what Algorithm 1 actually computes, and the experimental comparison conflates the information advantage with the equilibrium-seeking mechanism. If the authors are willing to substantially re-scope the claims—either by weakening the equilibrium terminology to 'intention-sharing MPC' or by adding a Nikaido-Isoda residual check and a control experiment—the paper could become publishable. I would not reject it outright because the experimental infrastructure and the reported trends are genuinely useful, but the current version is not suitable for publication without these changes."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper is worth knowing for the experimental result, not for the theory. They built a mixed-reality vehicle-in-the-loop testbed with a real Mazda CX-7 and three ghost vehicles, and measured 11.2–15.5% average fuel savings (up to 26.8%) for the ego vehicle in lane-change maneuvers when planners exchange over-the-horizon intentions over V2V, compared with a baseline MPC that predicts other vehicles unilaterally. That is a meaningful data point for the connected-vehicle energy literature. The testbed is a real engineering contribution.\n\nThe theoretical contribution is not new, and the paper mostly admits it: Lemma 1 and Theorem 1 restate the OCP-to-GNE result from Dreves and Gerdts [10]. That is fine if the packaging is honest, and it mostly is.\n\nThe load-bearing soft spot is the \"GNE\" label on the MPC trajectories. Algorithm 1 does one solve per agent per step using intentions broadcast before the current loop iteration. So each agent is responding to stale plans of others, not to the fixed point x*_¬v that Theorem 1 requires. The paper concedes closed-loop GNE cannot be claimed, but the issue is deeper: even the open-loop plans are not shown to satisfy the Nikaido-Isoda condition. The MIQP also has an integer lane command, so the convexity hypotheses of [10] are unverified. The energy savings may be real, but the paper has not shown they come from approaching a GNE rather than from simply receiving accurate future information.\n\nExperimental reporting is a bit thin: the ViL runs look like single trials without error bars, one sub-scenario needed post-hoc re-tuning, and one sub-scenario had a negative result (-12.2%). The baseline predictor parameters are partially reported, but the cost weights and some model constants are absent.\n\nStill, the central engineering claim — intention sharing reduces braking and saves fuel in a real vehicle — is plausible and supported by the trajectory and cost plots. The paper deserves a serious referee, but the theory section needs rewriting to claim only \"intention-sharing-based distributed MPC\", not GNE convergence. The experimental section should add trials or honest uncertainty bounds.\n\nRecommendation: send to peer review. It is a useful experimental paper with an over-claimed theoretical wrapper.\n\nTake care.","headline":"A useful vehicle-in-the-loop fuel-saving result wrapped in an over-claimed GNE theory; the experiments deserve review, the equilibrium label does not.","tokens_in":15523,"tokens_out":1778,"would_cite":true,"duration_ms":17381,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Connected vehicles that share their over-the-horizon plans can turn an interactive lane-change game into individual optimal-control problems, and a real vehicle running that scheme used 11.2 to 15.5 percent less fuel on average than a…","keywords":["generalized Nash equilibrium problem","connected automated vehicles","model predictive control","V2V intention sharing","lane change planning","vehicle-in-the-loop experiments","mixed-integer quadratic program","energy efficient driving"],"falsifier":"Run GNEP-MPC and the baseline in the same scenario while adding artificial delay or dropout to the V2V intention messages; if the energy savings disappear or reverse with realistic latency, the benefit depends on ideal communication. Separately, log the shared trajectories and check after each step whether each vehicle's plan is a best response to the others' final plans; if the joint state deviates from the mutual-best-response fixed point beyond tolerance, the open-loop generalized Nash equilibrium claim is not supported by the experiment.","tokens_in":14453,"feed_emoji":"🚗","tokens_out":6768,"duration_ms":61531,"temperature":0.7,"pith_summary":"The paper claims that vehicle-to-vehicle intention sharing converts a multi-agent lane-change motion planning problem, a generalized Nash equilibrium problem with coupled collision-avoidance constraints, into a set of independent optimal control problems, one per vehicle. That reduction makes distributed model predictive control possible, because each vehicle solves its own mixed-integer quadratic program against the shared trajectories of the others. The authors test the scheme in software-in-the-loop and then with one real vehicle embedded in a mixed-reality testbed, reporting 2.9 to 8.2 percent energy savings in software and up to 26.8 percent, with an 11.2 to 15.5 percent average, fuel savings in vehicle-in-the-loop tests relative to a baseline MPC that predicts other vehicles unilaterally. The stated physical mechanism is that accurate interactive predictions reduce braking and harsh corrective maneuvers during lane changes.","feed_headline":"Car-to-car intent sharing cuts fuel up to 26.8% in tests","feed_subtitle":"A real vehicle using shared-horizon plans as a distributed game solver burned 11.2 to 15.5 percent less fuel on average","key_machinery":"The enabling mechanism is Lemma 1's reduction: given $x^*_{\\neg v}$, the safe set $X_{\\mathrm{safe}}(x^k_{\\neg v})$ is fully determined over the horizon, so the coupled constraint (2c) becomes ordinary inequality constraints on the ego vehicle's own states and the game decouples into an optimal control problem per vehicle. The equilibrium check is the Nikaido-Isoda function, whose vanishing supremum characterizes a generalized Nash equilibrium; Theorem 1 uses it to argue that simultaneously solving the decoupled optimal control problems produces a generalized Nash equilibrium. The numerical machinery is a receding-horizon mixed-integer quadratic program with linear longitudinal and lateral models, a first-order powertrain lag, a second-order lateral response, integer lane commands, and big-M-encoded collision-avoidance logic implemented in the GNEP-MPC algorithm.","core_discovery":"The central claim is formal: if the optimal future state trajectories of the interacting vehicles, $x^*_{\\neg v}$, are available over the horizon, then each vehicle's dynamic generalized Nash equilibrium problem reduces to an ordinary optimal control problem, because the collision-avoidance safe set becomes a fixed set of constraints on the ego vehicle alone. Invoking the cited theorem that an optimal control solution for every agent is a generalized Nash equilibrium when costs depend only on own states, the paper concludes that the distributed MPC solutions are open-loop generalized Nash equilibria. The implemented planner is a mixed-integer quadratic program with an integer lane command, binary collision-avoidance variables, and receding-horizon execution; the experimental contribution is showing that these open-loop equilibrium-approximating plans outperform a strong baseline that predicts other traffic with an Intelligent Driver Model and chance constraints.","pith_inferences":["Delayed, quantized, or missing intention messages would break the exact reduction in Lemma 1; a natural test is to vary communication latency or dropout and measure how quickly the energy savings degrade.","Because Algorithm 1 performs a single solve per vehicle per step against possibly stale intentions, convergence to the fixed point $x^*_{\\neg v}$ is not established; running several best-response iterations per step and checking whether shared trajectories stabilize would test whether the generalized Nash equilibrium label is more than an approximation.","The formulation's costs depend only on own states, so the same Lemma 1 reduction should carry over to other interaction-intensive maneuvers such as merges, roundabouts, and intersection crossings; the empirical savings claims, however, are specific to the two-lane lane-change scenario tested."],"forward_implications":["Distributed planning becomes practical: each connected vehicle can compute its own plan locally as long as it receives the others' over-the-horizon trajectories, with no central coordinator.","Avoidance of an impeding vehicle no longer relies on guessing, and the reported fuel savings are physically tied to reduced braking and less frequent late lane changes.","The benefit appears in strong-interaction situations, where vehicles are close enough to affect each other's costs, and the game approach beats a unilateral-prediction baseline.","Receding-horizon execution of open-loop generalized Nash equilibrium plans is enough to improve closed-loop performance, even without claiming closed-loop convergence.","The mixed-reality vehicle-in-the-loop testbed can evaluate such planners with one real vehicle and simulated traffic, so the measured fuel savings reflect a real powertrain rather than a pure simulator."],"supporting_citations":[{"why":"Supplies the theorem that an optimal control solution for each agent is a generalized Nash equilibrium under convexity and compactness assumptions; the reduction's equilibrium conclusion rests on it.","marker":"[10]"},{"why":"Provides the definition of a generalized Nash equilibrium and the Nikaido-Isoda function used in the equilibrium argument.","marker":"[9]"},{"why":"Supplies the MIQP MPC formulation and the linear vehicle model that the GNEP-MPC planner is built on.","marker":"[18]"},{"why":"Supplies the Intelligent Driver Model used in the baseline to predict other vehicles' longitudinal motion.","marker":"[24]"},{"why":"Supplies the tractive-force fuel consumption model used to convert measured trajectories into energy figures.","marker":"[5]"},{"why":"Supplies the chance-constraint safety margin method used by the baseline prediction model.","marker":"[25]"},{"why":"Supplies the OBD-II fuel measurement approach used to compute the real vehicle's fuel consumption.","marker":"[26]"}],"fun_headline_variants":["Game-theoretic planner cuts fuel 12-15% in vehicle tests","Intent-sharing game solver saves fuel in real drive tests","GNEP planner reduces fuel use 12-15% in real experiments","Shared-horizon game planner burns 12-15% less fuel on road","Vehicle-in-loop game planner cuts fuel up to 27% in tests"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that each vehicle solves against the same fixed optimal state trajectories of the others over the whole horizon under the cited theorem's convexity and compactness assumptions; the implementation replaces that with a single receding-horizon solve per step against possibly stale intentions and an integer lane command, so the exact generalized Nash equilibrium claim is not what the hardware run verifies.","fun_headline_variants_meta":{"raw":{"variants":["Game-theoretic planner cuts fuel 12-15% in vehicle tests","Intent-sharing game solver saves fuel in real drive tests","GNEP planner reduces fuel use 12-15% in real experiments","Shared-horizon game planner burns 12-15% less fuel on road","Vehicle-in-loop game planner cuts fuel up to 27% in tests"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000883,"raw_usage":{"total_tokens":3767,"prompt_tokens":853,"completion_tokens":2914,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":469,"completion_tokens_details":{"reasoning_tokens":2820}},"tokens_in":469,"tokens_out":2914,"duration_ms":18509,"temperature":1.0,"reasoning_tokens":2820,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T15:08:50.821761+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run GNEP-MPC and the baseline in the same scenario while adding artificial delay or dropout to the V2V intention messages; if the energy savings disappear or reverse with realistic latency, the benefit depends on ideal communication. Separately, log the shared trajectories and check after each step whether each vehicle's plan is a best response to the others' final plans; if the joint state deviates from the mutual-best-response fixed point beyond tolerance, the open-loop generalized Nash equilibrium claim is not supported by the experiment.","supporting_citations":[{"cited_title":"Fundamentals of energy efcient driving for combustion engine and electric vehicles: An optimal control perspective,","cited_arxiv_id":null,"evidence_quote":"Supplies the tractive-force fuel consumption model used to convert measured trajectories into energy figures."},{"cited_title":"Generalized nash equilibrium problems,","cited_arxiv_id":null,"evidence_quote":"Provides the definition of a generalized Nash equilibrium and the Nikaido-Isoda function used in the equilibrium argument."},{"cited_title":"Multilane automated driving with optimal control and mixed-integer programming,","cited_arxiv_id":null,"evidence_quote":"Supplies the MIQP MPC formulation and the linear vehicle model that the GNEP-MPC planner is built on."},{"cited_title":"Enhanced intelligent driver model to access the impact of driving strategies on trafc capacity,","cited_arxiv_id":null,"evidence_quote":"Supplies the Intelligent Driver Model used in the baseline to predict other vehicles' longitudinal motion."},{"cited_title":"Automated vehicles in hazardous merging trafc: A chance-constrained approach,","cited_arxiv_id":null,"evidence_quote":"Supplies the chance-constraint safety margin method used by the baseline prediction model."},{"cited_title":"Energy andow effects of optimal automated driving in mixed trafc: V ehicle-in-the-loop experimental results,","cited_arxiv_id":null,"evidence_quote":"Supplies the OBD-II fuel measurement approach used to compute the real vehicle's fuel consumption."}],"review_version":1}