{"id":"27f08cdc-942d-4848-97d4-645999d168c2","arxiv_id":"2507.09822","paper_version":4,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":9,"one_line_summary":"An MPC framework that uses a Wasserstein-based risk metric and a Boltzmann model of agent behavior to actively probe and infer other vehicles' intentions in multimodal prediction settings.","lead":"This paper combines motion planning with multimodal trajectory predictions and active probing, where the car deliberately acts to reveal other drivers' intentions. The framework is tested in simulated merging and intersection scenarios, showing higher success rates than a baseline planner.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Lemma 1's closed-form Wasserstein risk (Eq. 8) has a sign error: for equal means/covariances it returns a negative squared distance, so the risk used in planning is not the proposed metric; the empirical claim is not tied to the stated framework until corrected.","rationale":"The reader's weakest assumption concerns the reward-model match between the ego's assumed behavior model and the simulated agents. That is a legitimate evaluation limitation, but the more decisive problem is that the core risk metric is mis-specified in the text. Eq. (8) is presented as a closed-form 2-Wasserstein distance, yet it has the wrong sign on all covariance trace terms and is not nonnegative. This is not a tuning or baseline issue: the safety cost in Eq. (9) is constructed from this risk, so the experiments cannot be reproduced from the paper as written. A single analytic plug-in settles the point, and if the sign is corrected the Monte Carlo results may survive, but until then the claimed superiority of active probing is not tied to the proposed framework. The verdict remains CONDITIONAL because the issue is fixable in revision, but the required corrections now include re-deriving Eq. (8) and re-running the experiments with the corrected risk, in addition to the reader's requested baselines and belief-update specification.","tokens_in":12426,"tokens_out":8914,"duration_ms":110237,"concrete_test":"Take one time step with \\bar{x}_e = \\hat{x}_i^k and C_e = C_i^k = I. The standard 2-Wasserstein squared distance is 0, while Eq. (8) evaluates to -2 Tr(3I) < 0. Then recompute the risk values entering Eq. (9) with the corrected Gaussian formula W_2^2 = ||\\mu_e - \\mu_i||^2 + Tr(C_e + C_i^k - 2(C_e^{1/2} C_i^k C_e^{1/2})^{1/2}) and rerun the Table II Monte Carlo. If the planned trajectories or success rates change materially, the paper's empirical claim does not correspond to its stated risk metric.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The most load-bearing concern is internal inconsistency in the risk computation that defines the planner, not just the evaluation setup. Lemma 1 presents the 2-Wasserstein distance as W_2^2 = ||\\bar{x}_e - \\hat{x}_i^k||^2 - 2 Tr[C_e + C_i^k + (C_e^{1/2} C_i^k C_e^{1/2})^{1/2}]. The standard Gaussian formula is W_2^2 = ||\\mu_e - \\mu_i||^2 + Tr(C_e + C_i^k - 2(C_e^{1/2} C_i^k C_e^{1/2})^{1/2}). The printed expression is not a distance: with \\bar{x}_e = \\hat{x}_i^k and C_e = C_i^k = I, Eq. (8) evaluates to -2 Tr(3I) < 0, whereas the true squared distance is 0. Hence the quantity entering Eq. (7) via exp(-\\alpha W) is not a well-defined real risk, and the claim that the risk is 'always finite' is beside the point: the formula is not the object it claims to be. Since Eq. (9) builds the safety objective from this risk, the planner evaluated in Tables II and III is not completely specified by the text. The headline result (98% vs 62% success) therefore cannot be attributed to the stated algorithm without either correcting Eq. (8) or releasing the exact cost computation used in the experiments.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a motion-planning framework that integrates MPC with multimodal trajectory predictions and an active-probing mechanism. The ego agent optimizes a weighted sum of utility, a safety reward based on a Wasserstein-type risk over predicted Gaussian mixture modes, and an information-gain objective that estimates other agents' behavior parameters through a Boltzmann likelihood and a particle filter. The framework is evaluated in MetaDrive in a lane-change scenario and an unsignalized-intersection scenario. The lane-change experiments generate aggressive/defensive agents from a reward model whose feature structure matches the ego's assumed model, while the intersection experiments use MetaDrive's independent RL controllers. Tables II and III report higher success rates and faster merge/cross times for the active-probing planner compared with a chance-constrained MPC baseline and a no-probing variant.","tokens_in":12732,"tokens_out":6431,"duration_ms":77142,"significance":"If the reported results hold, the paper demonstrates a practically relevant benefit of active probing: 98% versus 62% lane-change success and 96% versus 89% intersection success, with zero intersection collisions, alongside released code and videos. The intersection scenario with independent RL controllers is a valuable generality check that goes beyond an in-model-class evaluation. However, the central planner specification currently contains a load-bearing error in the risk formula, and the information-gain update is underspecified, so the experimental numbers cannot yet be attributed to the stated algorithm without correction.","major_comments":[{"comment":"The printed closed-form expression cannot be a 2-Wasserstein distance. For equal means and identity covariances, Eq. (8) evaluates to -2 Tr(3I) < 0 instead of 0, whereas the true squared 2-Wasserstein distance is 0. The standard Gaussian formula is W_2^2 = ||\\bar{x}_e - \\hat{x}_i^k||^2 + Tr(C_e + C_i - 2(C_e^{1/2} C_i C_e^{1/2})^{1/2}). Since Eq. (7) and the safety cost (9) are built on this quantity, the planner whose results appear in Tables II and III is not fully specified. Please correct Eq. (8) and confirm that the experimental implementation uses the corrected expression; the 'always finite' statement in Lemma 1 does not address the sign error.","section":"III-B, Eq. (8)"},{"comment":"The particle-filter measurement update is under-specified. The posterior b'_k is written only as proportional to b(\\phi) \\hat{p}^i_k(x_e, u_e), but \\hat{p}^i_k depends on \\phi through R_i in Eq. (10); the exact likelihood function, its normalization, the resampling step, and the precise KL estimator over M particles are not given. Without these details, the information-gain objective (12) cannot be reproduced from the text, and it is unclear whether the reported probing benefit comes from the stated belief update.","section":"III-C, Eqs. (11)-(12) and Algorithm 1"},{"comment":"The quoted risk threshold \\tau = 5 can never trigger, because r^i_k(t) = p^i_k(1 + e^{-\\alpha W}) \\le 2 for all t when p^i_k \\le 1 and the exponential term is at most 1. Thus the safety cutoff on probing is inactive in every reported experiment, and any observed benefit of probing cannot be attributed to the threshold mechanism. Please either report the actual threshold used or modify the risk scaling so that the threshold is meaningful.","section":"Table I and Eq. (12)"},{"comment":"The sentence 'our algorithm with probing disabled (\\alpha_2 = 0)' is inconsistent with the objective (2), in which \\alpha_2 weights the safety term and \\alpha_3 weights the information term. If the baseline truly set \\alpha_2 = 0, it would have no safety objective; if the intended weight was \\alpha_3 = 0, the text should say so. This matters because the 'No probing' row in Table II is one of the two baselines used to attribute the observed improvement to probing.","section":"IV-C, Table II"},{"comment":"The lane-change controllers are generated by optimizing a reward function whose features (velocity matching, inter-agent safety distance, and lane deviation) match the feature structure that the ego assumes in its reward model (15). These experiments therefore primarily validate behavior estimation within the assumed model class. The intersection experiment with MetaDrive RL controllers provides important out-of-model evidence, but the paper should state this limitation explicitly and, if possible, include a lane-change condition with misspecified features.","section":"IV-A and Eqs. (15)-(16)"}],"minor_comments":[{"comment":"The proof's 'state space is bounded' justification is inconsistent with Gaussian distributions over an unbounded state space; finiteness of the risk follows from finite means and covariances, so the proof should be reworded.","section":"III-B, Lemma 1"},{"comment":"The table omits the risk sensitivity \\alpha from Eq. (7), the prior covariance \\Sigma, and the particle count M used in the belief update; these values are needed to reproduce the experiments.","section":"Table I"},{"comment":"The threshold check uses r^i_k(t) without a loop over t, so it is unclear whether the cutoff applies if the risk exceeds \\tau at any time step, at every time step, or at a particular time; please specify the intended condition.","section":"Algorithm 1"},{"comment":"The time dependence of Info is ambiguous: the information gain is written without a time index, but the cutoff condition depends on r^i_k(t); please clarify how the per-time-step risk feeds into the information objective.","section":"Eq. (12)"}],"recommendation":"major_revision","confidential_remarks":"The Eq. (8) sign error is likely typographical, but because the planner in the experiments is not fully specified, I could not verify that the reported numbers use a correctly defined risk. I recommend requiring the authors to state the exact implementation formula and to provide code or parameter values for the particle filter. The intersection results give genuine out-of-model evidence, so I would not reject on circularity grounds; the issues appear fixable within the manuscript's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the framework is genuinely interesting, but the paper cannot be trusted in its current form because Lemma 1's closed-form Wasserstein formula is wrong. The stress-test is correct: Eq. (8) gives negative values when means and covariances match. The risk that enters the planner is therefore not the risk they define. This is not a cosmetic typo; it is load-bearing because the safety objective in Eq. (9) is built from that risk. That said, the empirical work shows the idea can work, and the intersection experiment with MetaDrive's built-in RL controllers is the strongest part because it does not share the ego's reward model.\n\nWhat is new? The risk metric p_i^k(1 + exp(-alpha W)) is a reasonable way to blend mode likelihood with distributional distance. The modified Boltzmann model that scales mode rewards by the predictor's confidence is a sensible touch. The integration into MPC with an information-gain objective is coherent.\n\nWhere are the soft spots? The sign error is the big one. The lane-change experiments are partly self-fulfilling because the simulated agents are generated from the same reward feature vector the ego assumes. No error bars appear on the Monte Carlo tables. The threshold tau = 5 is vacuous since the risk is bounded well below 5. The particle filter update is described only loosely, and the Monte Carlo KL estimate appears to be misweighted. All of these are fixable, but they make the empirical claims hard to verify.\n\nBottom line: this is a paper for the motion planning subfield, with a real idea and a plausible demonstration. But it needs a major revision before publication. I would send it to reviewers, not desk-reject it.","headline":"Active probing is a worthwhile idea, but the printed Wasserstein formula in Lemma 1 is wrong; the paper needs a corrective revision before the empirical claims can be trusted.","tokens_in":13283,"tokens_out":3236,"would_cite":false,"duration_ms":36141,"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":"A planner that actively probes other drivers merges successfully 98% of the time, versus 62% for a waiting planner.","keywords":["motion planning","active probing","multimodal prediction","risk metric","Wasserstein distance","behavior inference","particle filter","Model Predictive Control"],"falsifier":"Run the lane-change Monte Carlo with true agents that optimize a reward containing a feature absent from Eq. (15), for example a politeness or gap-acceptance term, and measure whether probing still beats the no-probing variant on success rate and merge time; if the advantage collapses, the probing claim depends on the assumed reward family.","tokens_in":12205,"feed_emoji":"🚗","tokens_out":7418,"duration_ms":82998,"temperature":0.7,"pith_summary":"An autonomous planner can deliberately choose gentle accelerating or steering probes to make surrounding drivers reveal whether they are yielding or aggressive, and folding that information into a multimodal risk-aware objective produces safer, faster merging and intersection crossing than passively waiting. The paper builds one objective out of a utility cost, a Wasserstein-based safety risk that has a closed form when prediction modes are Gaussian, and an information-gain term that updates a particle-filter belief over each agent's reward weights through a Boltzmann model that depends on the ego's own actions. In MetaDrive simulations, the active probing variant completes the lane change in 98% of 200 Monte Carlo episodes against 62% for the chance-constrained MPC baseline, and crosses an unsignalized intersection in 96% of episodes against 89%, with zero collisions. The reason to care is that the result recasts cautious driving as an information-gathering problem: the ego does not have to wait for intentions to be revealed if it can nudge them out.","feed_headline":"Active probing lifts merge success to 98%","feed_subtitle":"Nudging other cars reveals whether they will yield, letting the ego merge faster than a chance-constrained planner that waits.","key_machinery":"The load-bearing object is the modified Boltzmann likelihood of Eq. (11), which ties the probability of each predicted mode $k$ for agent $i$ to the reward that mode would give agent $i$ under the ego's candidate action $u_e$. Because the reward model $R^i(x^i,u^i,s^i(x_e,u_e))$ depends on the ego's trajectory through the interaction set $s^i$, the planner's information-gain objective becomes a function of its own controls, so gradient-based MPC can select probing actions. The second essential piece is Lemma 1: with Gaussian mode distributions, the 2-Wasserstein distance in the risk metric is finite and explicit, which keeps the safety cost tractable and lets a risk threshold $\\tau$ switch probing off in dangerous modes.","core_discovery":"The central claim is that active probing with multimodal predictions is both tractable and effective: the ego vehicle can estimate each other agent's behavior parameters $\\phi^i$ online and use the estimate to plan. The paper defines a risk $r_i^k(t) = p_i^k(1 + e^{-\\alpha W(P^e(t), P_i^k(t))})$ over each predicted mode, with $W$ the 2-Wasserstein distance, and proves (Lemma 1) that when the ego and mode distributions are Gaussian, the risk is always finite and computable in closed form. It then makes the predictor's mode likelihoods action-dependent through a modified Boltzmann model $\\hat{p}_i^k(x_e,u_e) \\propto p_i^k e^{R^i}$, so that the ego's choice of control changes the expected information gain about $\\phi^i$. The experiments support the claim in the paper's own comparison: 98% versus 62% lane-change success, 96% versus 89% intersection success, and faster merge and crossing times than the chance-constrained planner.","pith_inferences":["The lane-change gain may depend on feature alignment: the paper samples true aggressive and defensive controllers from the same reward features the ego assumes, so a test with qualitatively different driver features would show how much of the 98% is probing skill versus model-matching.","A natural extension is to make the probe gain depend on prediction confidence: when the predictor's mode likelihoods $p_i^k$ are miscalibrated, the Boltzmann rescaling could amplify noise, so a sensitivity study over miscalibrated likelihoods would sharpen the practical claims.","The 2.5-second planning horizon and 0.1-second discretization restrict probes to short nudges; real highway merging often requires seconds-long interactions, so a longer-horizon variant would test whether the mechanism still works when the reveal signal arrives later.","The risk threshold $\\tau$ currently gates probing per mode; an adaptive threshold that raises when traffic density or prediction covariance grows could make the same idea safer in denser scenes."],"forward_implications":["Active probing removes the need to wait passively: in the paper's lane-change Monte Carlo, merging success rises from 62% with chance-constrained MPC to 98% with probing.","Probing also shortens the maneuver: reported mean time to merge drops from 9.27 seconds for the baseline to 6.87 seconds with probing.","The closed-form finite risk metric means the safety objective can be evaluated analytically for Gaussian mixture predictions rather than by sampling trajectories.","Because interactions are captured through per-agent reward models instead of a joint dynamics model, adding prediction modes does not blow up the optimization's state space.","The framework carries over to agents it did not train on: in the intersection scenario with MetaDrive's built-in controllers, probing still achieves the highest success rate, 96%, and zero collisions."],"supporting_citations":[{"why":"Supplies the chance-constrained MPC baseline whose 62% and 89% success rates the probing results are compared against.","marker":"[19]"},{"why":"Source of the information-radius and active-probing objective and the Boltzmann-style choice model that Eq. (11) modifies.","marker":"[27]"},{"why":"The Gaussian Lane Keeping predictor that generates the K=3 trajectories, likelihoods, and covariances used in all simulations.","marker":"[35]"},{"why":"MetaDrive simulator provides the driving scenarios, vehicle dynamics, and the built-in intersection agents that test model mismatch.","marker":"[10]"},{"why":"The Intelligent Driver Model inspires the reward features (velocity matching, safe distance, lane keeping) in Eq. (15).","marker":"[36]"},{"why":"Provides the definition and properties of the 2-Wasserstein distance used in the risk metric and its finiteness argument.","marker":"[28]"},{"why":"Supports the Boltzmann model of agent decision-making behind the modified likelihood update.","marker":"[32]"}],"fun_headline_variants":["Probing other vehicles yields 98% merge success in tests","Closed-form risk model enables active probing for planners","Active probing cuts merge time, hits 98% success","Probing other drivers reveals yield likelihood for safer merges","Active probing boosts lane-change to 98% from 62%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole probing scheme assumes that other drivers choose actions to maximize the ego's specified reward features, namely velocity matching, safe distance, and lane keeping, so that learning the feature weights tells the ego what they will do next.","fun_headline_variants_meta":{"raw":{"variants":["Probing other vehicles yields 98% merge success in tests","Closed-form risk model enables active probing for planners","Active probing cuts merge time, hits 98% success","Probing other drivers reveals yield likelihood for safer merges","Active probing boosts lane-change to 98% from 62%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000892,"raw_usage":{"total_tokens":3844,"prompt_tokens":940,"completion_tokens":2904,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":556,"completion_tokens_details":{"reasoning_tokens":2822}},"tokens_in":556,"tokens_out":2904,"duration_ms":23769,"temperature":1.0,"reasoning_tokens":2822,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T17:46:50.583483+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the lane-change Monte Carlo with true agents that optimize a reward containing a feature absent from Eq. (15), for example a politeness or gap-acceptance term, and measure whether probing still beats the no-probing variant on success rate and merge time; if the advantage collapses, the probing claim depends on the assumed reward family.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the chance-constrained MPC baseline whose 62% and 89% success rates the probing results are compared against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Source of the information-radius and active-probing objective and the Boltzmann-style choice model that Eq. (11) modifies."},{"cited_title":"Gaussian Lane Keeping: A Robust Prediction Baseline","cited_arxiv_id":"2407.18451","evidence_quote":"The Gaussian Lane Keeping predictor that generates the K=3 trajectories, likelihoods, and covariances used in all simulations."},{"cited_title":"Treiber, A","cited_arxiv_id":null,"evidence_quote":"The Intelligent Driver Model inspires the reward features (velocity matching, safe distance, lane keeping) in Eq. (15)."},{"cited_title":"Peyr ´e, M","cited_arxiv_id":null,"evidence_quote":"Provides the definition and properties of the 2-Wasserstein distance used in the risk metric and its finiteness argument."}],"review_version":1}