{"id":"ffe758be-fc26-4733-9bdf-579e5368c4ba","arxiv_id":"2507.20850","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"A free-energy-inspired cognitive risk signal adapts pedestrian forces and a GCN-SAC vehicle controller, improving safety and efficiency in a three-pedestrian shared-space simulation.","lead":"This paper connects a free-energy-inspired cognitive uncertainty measure to pedestrian movement modeling and to a graph-based reinforcement learning controller for autonomous vehicles. The combined framework reports safer, faster, and smoother simulated navigation in a three-pedestrian shared-space dataset.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Cognitive uncertainty is not well-defined: Algorithm 1 requires an 'internal prediction model' that the paper never specifies, and every risk weight in CR-SFM plus the GCN adjacency in G-SAC-Cog inherits this undefined quantity, so the reported gains cannot be attributed to the FEP-inspired…","rationale":"This is a simulation-only method paper, but the SAC/GCN integration is internally coherent and the risk-weighted social force equations are self-consistent. The paper deserves credit for explicit ablations (RA-SFM, G-SAC-NoCog) and for using real HBS initial conditions. However, the strongest claim rests on u_t, and u_t is undefined because the internal prediction model is never formalized. This is the same weakest assumption the reader identified, and it is load-bearing: every risk modulation in both the pedestrian model and the AV graph either inherits u_t or disappears when u_t is removed. The statistical concerns (point estimates without error bars over roughly 18 test episodes) amplify the issue but are secondary. The qualitative claim in Fig. 3c that CR-SFM proactively detours 'even when physical risk is low' is precisely the phenomenon controlled by the hidden predictor. My recommendation is not to lower the reader's CONDITIONAL verdict further: the core architecture is plausible, and the concern is addressable by disclosure and a sensitivity analysis. The required conditions are to specify the internal prediction model, release code and hyperparameters, and report repeated-seed statistics. If those are satisfied, the claim can be evaluated on its merits; until then, the central mechanism is not reproducible.","tokens_in":18010,"tokens_out":4714,"duration_ms":50796,"concrete_test":"Ask the authors to specify the internal prediction model in Algorithm 1, including how \\v_i^t sets µ_p and σ_p^2 and how the previous posterior is combined. Then rerun the full pipeline with three concrete instantiations: (i) constant-velocity predictor \\v_i^t = v_i^{t−1}; (ii) posterior-mean predictor \\v_i^t = µ_post,t−1; (iii) oracle predictor \\v_i^t = v_i^t, so u_t ≈ 0. Recompute Table I (CR-SFM ADE/FDE/CR) and Table II (G-SAC-Cog success/collision/speed/jerk) for each. If the reported 0.94 success rate, 0.06 collision rate, and 0.0 pedestrian collision rate move outside the sampling uncertainty of the original runs, or if the oracle variant still yields the same gains, the cognitive-uncertainty mechanism is not the operative cause and the central claim fails as stated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central mechanism is Eq. (5), u_t = D_KL(p_t(v)||o_t(v)), claimed to quantify cognitive uncertainty. Computing it requires the predicted velocity \\h\\v_i^t that defines p_t(v), but Algorithm 1 (Section IV-A) merely says 'Use internal prediction model to predict velocity \\v_i^t'; no model, equations, or parameter values are provided, and no guidance is given for how \\v_i^t combines with the previous posterior p_post,t−1(v) to set µ_p and σ_p^2. This is not a cosmetic omission: u_t enters Eqs. (16)–(18) as w_veh, w_ped, and w_goal in CR-SFM, and Eq. (25) as the adjacency matrix A_t(i,j) for the GCN in G-SAC-Cog. The entire 'cognitive' contribution of the paper is therefore determined by an unspecified component. Different plausible choices—constant-velocity extrapolation, a learned social-force predictor, or an oracle using the observed v_i^t (which would drive u_t toward zero if σ_o is fixed)—would in general produce different force weights and graph edges, and hence different Table I and Table II numbers. The qualitative claims in Section VI that CR-SFM 'detects latent threats' (Fig. 3c) and that G-SAC-Cog avoids collisions better are exactly the effects that depend on this hidden predictor. Without specifying it, the results are not reproducible, and the improvement cannot be attributed to FEP-inspired uncertainty rather than to an implicit oracle or arbitrary calibration. The RA-SFM versus CR-SFM ablation is the cleanest evidence for the cognitive term, but it is invalidated as evidence until the prediction model is pinned down.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a framework for AV navigation in pedestrian-rich shared spaces, built around a Free-Energy-Principle-style measure of cognitive uncertainty. The uncertainty is defined in Eq. (5) as the KL divergence between a predicted and an observed Gaussian velocity distribution, and it is used in two places: the pedestrian Cognitive-Risk Social Force Model (CR-SFM), where it modulates goal and repulsive force weights (Eqs. 12-18), and the AV decision-making model G-SAC-Cog, where it weights a dynamic adjacency matrix for a GCN integrated into Soft Actor-Critic (Eq. 25). Experiments are carried out on scenarios extracted from the HBS dataset. Table I reports ADE/FDE/collision rate for CR-SFM against CV, SFM, and RA-SFM, and Table II reports success rate, collision rate, speed, and jerk for G-SAC-Cog against UAW-PCG, S-SAC, G-SAC-NoCog, and a human driver. The central claim is that the cognitive-uncertainty signal improves both pedestrian trajectory realism and AV safety/efficiency/smoothness.","tokens_in":18454,"tokens_out":3698,"duration_ms":44689,"significance":"If the mechanism were fully specified, the paper would address a real gap: interactive, responsive pedestrian simulation for AV decision-making in shared spaces, with a risk-aware graph representation inside a state-of-the-art RL algorithm. The combination of a social-force pedestrian model with a GCN-SAC controller is sensible, and the qualitative examples in Figs. 4-5 are informative. However, the current manuscript cannot support its central attribution claim, because the core cognitive-uncertainty signal is produced by an internal prediction model that is never defined, and because the experimental reporting relies on single-point metrics without variance or a clear train/test separation. The claimed gains therefore cannot be reproduced or traced to the proposed FEP-inspired mechanism as written.","major_comments":[{"comment":"The internal prediction model is listed as an input to Algorithm 1 but is never specified. The predicted velocity \\hat{v}_i^t determines the prior p_t(v), and therefore the cognitive uncertainty u_t in Eq. (5). This u_t enters every risk weight in CR-SFM (Eqs. 16-18) and the GCN adjacency matrix in G-SAC-Cog (Eq. 25). Different plausible choices, such as constant-velocity extrapolation, a learned social-force predictor, or an oracle using the observed velocity, will in general produce different force weights and graph edges, and hence different numbers in Tables I and II. The paper must specify the internal prediction model in full, including how it is initialized, updated, and combined with the posterior from the previous step, and should report sensitivity to this choice.","section":"Section IV-A, Algorithm 1, Eq. (5)"},{"comment":"The pedestrian model parameters are calibrated using Bayesian optimization on trajectories extracted from the training dataset, but the paper does not state whether the ADE/FDE/CR values in Table I are computed on a disjoint test set. If the reported values are on the calibration set, they partly reflect fitting quality rather than predictive fidelity, undermining the comparison between CR-SFM and the baselines. The authors should clearly specify the train/test split, the number of scenarios used for evaluation, and whether the reported metrics are averaged over that held-out set.","section":"Section V-B, Table I"},{"comment":"All AV performance numbers are single-point estimates with no standard deviations, confidence intervals, number of evaluation episodes, or multiple seeds. Given that the difference between G-SAC-Cog (success 0.94, collision 0.06) and UAW-PCG (success 0.83, collision 0.17) is based on what appears to be a limited number of HBS interaction scenarios, the reported superiority is not statistically supported. The authors should report per-episode results across several random seeds and provide significance measures or confidence intervals.","section":"Section VI-B, Table II"},{"comment":"The KL divergence D_KL(N(\\mu_p,\\sigma_p^2)||N(\\mu_o,\\sigma_o^2)) has a closed form in which, for fixed variances, u_t is a monotone function of the squared prediction error (\\mu_p - \\mu_o)^2. It is therefore unclear whether the proposed quantity captures anything beyond the predictor's raw error, and the FEP-inspired Bayesian update in Eqs. (6)-(7) does not by itself add information. The authors should state whether this is intended, and should justify why this particular proxy constitutes 'cognitive uncertainty' rather than simply prediction error, especially since the qualitative claim in Section VI-A that CR-SFM 'detects latent threats' relies entirely on this quantity.","section":"Section IV-A, Eq. (5)"},{"comment":"The hyperparameters \\gamma_1, \\gamma_2, \\lambda_1, \\lambda_2, \\lambda_3, \\lambda, \\sigma_o, and \\sigma_p are not reported anywhere in the manuscript. All risk weights and adjacency values depend on these parameters, and without them the experiments are not reproducible. The authors should provide the full parameter set used, including the ranges and fitted values from the Bayesian optimization procedure for the pedestrian models.","section":"Section V-A and Section IV"}],"minor_comments":[{"comment":"The text says 'FED' where it should say 'FDE' in the sentence describing SFM results; this typo should be corrected.","section":"Section VI-A.2"},{"comment":"There is a typo in 'statest' that should read 'state'.","section":"Section III"},{"comment":"Algorithm 1 says 'Use internal prediction model to predict velocity', but the phrase 'internal prediction model' is not defined in the algorithm or in the surrounding text; this is a clarity issue even beyond the substantive concern in Major Comment 1.","section":"Section IV-A, Algorithm 1"},{"comment":"The row labeled 'Human driver' reports '1' for success and collision rates with no indication of how many human-driving episodes were recorded; this should be stated for the comparison to be interpretable.","section":"Section V-B and Table II"},{"comment":"Reference [38] duplicates reference [15] (the same Helbing and Molnar social force paper); the duplicate should be removed or converted into a proper repeated citation.","section":"References"},{"comment":"The definition of \\phi in Eq. (13) is stated in prose as 'the angle between the agent's velocity vector and the vector pointing from the agent to pedestrian i', but it would help to draw the geometric configuration or provide an explicit vector formula, since the sign of cos\\phi affects the risk computation.","section":"Section IV-B, Eq. (13)"}],"recommendation":"major_revision","confidential_remarks":"The core issue is not that the FEP-inspired framework is conceptually wrong; it is that the implementation omits the central prediction mechanism. If the authors fully specify the internal prediction model, report hyperparameters, and provide statistically grounded evaluation, the paper could become a useful contribution to AV-pedestrian interaction modeling. The omission of the predictor is the main blocker and should be treated as a mandatory revision requirement."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper does one thing genuinely well—it embeds the same risk signal in both the pedestrian model and the AV policy, which gives a clean ablation story. The RA-SFM vs CR-SFM and G-SAC-NoCog vs G-SAC-Cog comparisons are the right way to isolate the contribution. But the key mechanism, the KL-divergence cognitive uncertainty u_t, is not actually defined. Algorithm 1 calls an 'internal prediction model' and never specifies it. Since u_t enters every risk weight (Eqs. 16–18) and the GCN adjacency (Eq. 25), the entire FEP-inspired component is a placeholder. Different predictors—constant velocity, a learned SFM, or an oracle—would change the numbers, so the reported improvements cannot be attributed to the cognitive term.\n\nThe math that is shown (Eqs. 2–7, 13–18, 24–25) is standard and the integration itself is coherent. The writing is clear and the qualitative figures do illustrate the intended behavior. Credit where it's due: the virtual-distance risk measure and the exponential goal-weight decay are sensible choices, and the choice to use the same fused risk in both models is a legitimate design decision.\n\nThe soft spots beyond the missing predictor: the evaluation uses only 78 scenarios (about 18 test episodes), reports single-point metrics without error bars, and no code or hyperparameters are released. The ADE/FDE calibration on the training split is reasonable but the test size makes the 0.78/0.44 numbers fragile. The AV results are from training and evaluating inside the same fitted simulator, so they are self-consistent but not validated against real interaction dynamics. These are common limitations in this subfield, and I'd expect a serious referee to request more thorough statistical reporting.\n\nThe undefined predictor is the load-bearing flaw. If the authors specify a concrete model (even a constant-velocity baseline) and show that the results are robust to that choice, the paper becomes a solid subfield contribution. As written, I would not cite it, but I would send it to review because the question is important and the fix is likely within reach.","headline":"A well-structured integration of risk-weighted social forces and graph-based SAC, but the cognitive-uncertainty mechanism depends on an unspecified predictor, so the reported gains are currently uninterpretable.","tokens_in":18937,"tokens_out":3137,"would_cite":false,"duration_ms":34553,"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":"This paper sets out to establish that a single cognitive-risk signal, derived from the Free Energy Principle, should govern both pedestrian reactions and an autonomous vehicle's policy in shared spaces.","keywords":["Autonomous vehicles","Pedestrian behavior modeling","Free energy principle","Cognitive uncertainty","Social force model","Graph neural networks","Soft actor-critic","Shared space navigation"],"falsifier":"Replace the undefined internal prediction model in Algorithm 1 with a trivial constant-velocity predictor and rerun the full simulation battery; if G-SAC-Cog and CR-SFM keep their margins over the ablations, the reported safety gain does not depend on the free-energy-inspired mechanism, whereas if the margins collapse, the gain is tied to that specific predictor.","tokens_in":17834,"feed_emoji":"🚗","tokens_out":5132,"duration_ms":55673,"temperature":0.7,"pith_summary":"The paper sets out to establish that a single cognitive-risk signal, derived from the Free Energy Principle, can make both simulated pedestrians and an autonomous vehicle's policy behave more realistically and safely in shared spaces. It proposes the Cognitive-Risk Social Force Model (CR-SFM) for pedestrians and a graph-enhanced Soft Actor-Critic agent (G-SAC-Cog) for the vehicle, with both sides using the same fused measure of physical risk and cognitive uncertainty. In simulations built on the Hamburg Bergedorf Station dataset, the full framework reaches its destination in 0.94 of episodes, collides in only 0.06, averages 4.470 m/s, and keeps jerk at 0.524 m/s³, beating the UAW-PCG baseline and both ablations. If the claim is right, cognitive uncertainty is a usable shared currency for modeling bidirectional human-vehicle interaction, not just an extra reward term.","feed_headline":"Free-energy risk signal lifts AV success to 94 percent","feed_subtitle":"One fused risk measure drives pedestrian reactions and the AV policy, cutting collisions to 6 percent in simulation.","key_machinery":"The load-bearing object is the fused risk measure: the physical risk ψ computed from a virtual distance that inflates or shrinks the geometric distance based on an agent's velocity, acceleration, and approach direction, multiplied by the cognitive term $(1 + \\lambda u)$, where $u$ is the KL-divergence uncertainty from the predict–observe–update loop. This product appears twice: as the force-weight coefficients $w^{\\text{goal}}$, $w^{\\text{veh}}$, $w^{\\text{ped}}$ in CR-SFM (Eqs. 16–18), and as the adjacency entry $A_t(i,j) = \\psi_t(i,j)(1 + \\lambda u_t(i,j))$ in the GCN (Eq. 25). The argument is that routing this one scalar signal through both the pedestrian model and the AV policy is what produces adaptive, bidirectional interaction rather than fixed-rule behavior.","core_discovery":"The central claim is that cognitive uncertainty — quantified as the KL divergence between an agent's predicted and observed velocity distributions — can be fused with physical risk into a single weight that makes both pedestrian trajectories and AV policies safer and more human-like. For pedestrians, the Cognitive-Risk Social Force Model multiplies goal, vehicle, and pedestrian repulsion forces by risk-based weights, so a nearby but receding agent produces little repulsion while an uncertain approaching one pulls the pedestrian off course. For the AV, the same fused risk defines the dynamic adjacency matrix of a GCN embedded in SAC's actor and critic, so the policy focuses its graph aggregation on high-risk agents. The quantitative claim is that G-SAC-Cog reaches a 0.94 success rate and 0.06 collision rate in the simulated shared-space scenarios, with the CR-SFM pedestrian model recording zero collisions and lower displacement errors than the SFM and RA-SFM baselines.","pith_inferences":["The paper's dependence on an unspecified 'internal prediction model' in Algorithm 1 means the reported gains could partly reflect that choice; a controlled swap of predictor (constant-velocity versus learned) would reveal how much of the effect is cognitive uncertainty rather than prediction quality.","The same adjacency-matrix design could be extended to variable pedestrian counts by letting nodes enter and leave the graph online, a setting the paper lists as future work but does not test.","Because the risk signal is shared between pedestrian and vehicle, the framework offers a direct way to inject the same uncertainty model into real-world traffic simulators or digital twins for scenario-based testing."],"forward_implications":["The CR-SFM pedestrian model produces zero collisions and lower ADE and FDE than SFM and RA-SFM, showing that the fused-risk weighting makes simulated pedestrians both more accurate and safer.","G-SAC-Cog outperforms UAW-PCG and both ablations on success rate, collision rate, and average speed, indicating that a risk-encoded adjacency matrix helps the policy perceive which interactions matter.","The jerk and maximum-acceleration results show that acceleration-based control with risk-weighted graph features yields smoother motion than velocity-based baselines, so safety does not come at the cost of comfort.","If the framework is correct, the same cognitive-risk signal can serve as a common interface between pedestrian simulation and AV decision-making, removing the need for fixed-behavior pedestrian models in training."],"supporting_citations":[{"why":"Supplies the real-world shared-space scenarios and ground-truth trajectories used for training and evaluation.","marker":"[43]"},{"why":"The state-of-the-art DRL baseline it must beat, using uncertainty-aware polar collision grids.","marker":"[10]"},{"why":"Provides the three-force pedestrian model that CR-SFM extends with risk-based weights.","marker":"[39]"},{"why":"Provides the off-policy maximum-entropy actor-critic algorithm on which the AV policy is built.","marker":"[41]"},{"why":"Supplies the graph convolutional propagation rule into which the risk-aware adjacency matrix is plugged.","marker":"[42]"},{"why":"The constant-velocity pedestrian baseline that shows what ignoring interaction costs.","marker":"[9]"},{"why":"Grounds the Free Energy Principle used to justify the predict–observe–update loop for cognitive uncertainty.","marker":"[35]"}],"fun_headline_variants":["Free-energy risk fusion lifts AV success to 94%","Cognitive risk boosts AV safety in pedestrian-rich scenes","AVs mimic human cognition for safer crowd navigation","Unified risk signal helps AVs navigate with pedestrians","Risk-aware GCN and social force raise AV success to 94%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The cognitive uncertainty $u_t$ is computed from a predicted velocity provided by an 'internal prediction model' that is never defined in the paper; if this predictor is arbitrary or unrealistic, the KL-divergence signal and every risk weight built on it lose their grounding.","fun_headline_variants_meta":{"raw":{"variants":["Free-energy risk fusion lifts AV success to 94%","Cognitive risk boosts AV safety in pedestrian-rich scenes","AVs mimic human cognition for safer crowd navigation","Unified risk signal helps AVs navigate with pedestrians","Risk-aware GCN and social force raise AV success to 94%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000731,"raw_usage":{"total_tokens":3280,"prompt_tokens":959,"completion_tokens":2321,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":575,"completion_tokens_details":{"reasoning_tokens":2242}},"tokens_in":575,"tokens_out":2321,"duration_ms":19825,"temperature":1.0,"reasoning_tokens":2242,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T13:11:29.978078+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Replace the undefined internal prediction model in Algorithm 1 with a trivial constant-velocity predictor and rerun the full simulation battery; if G-SAC-Cog and CR-SFM keep their margins over the ablations, the reported safety gain does not depend on the free-energy-inspired mechanism, whereas if the margins collapse, the gain is tied to that specific predictor.","supporting_citations":[{"cited_title":"A discrete choice model for solving conflict situations between pedestrians and vehicles in shared space","cited_arxiv_id":"1709.09412","evidence_quote":"Supplies the real-world shared-space scenarios and ground-truth trajectories used for training and evaluation."},{"cited_title":"Uncertainty-aware drl for autonomous vehicle crowd navigation in shared space,","cited_arxiv_id":null,"evidence_quote":"The state-of-the-art DRL baseline it must beat, using uncertainty-aware polar collision grids."},{"cited_title":"Interaction-aware decision-making for automated vehicles using social value orientation,","cited_arxiv_id":null,"evidence_quote":"Provides the three-force pedestrian model that CR-SFM extends with risk-based weights."},{"cited_title":"Soft actor-critic: Off- policy maximum entropy deep reinforcement learning with a stochastic actor,","cited_arxiv_id":null,"evidence_quote":"Provides the off-policy maximum-entropy actor-critic algorithm on which the AV policy is built."},{"cited_title":"Semi-supervised learning with graph learning-convolutional networks,","cited_arxiv_id":null,"evidence_quote":"Supplies the graph convolutional propagation rule into which the risk-aware adjacency matrix is plugged."},{"cited_title":"What the constant velocity model can teach us about pedestrian motion prediction,","cited_arxiv_id":null,"evidence_quote":"The constant-velocity pedestrian baseline that shows what ignoring interaction costs."},{"cited_title":"The free-energy principle: a unified brain theory?","cited_arxiv_id":null,"evidence_quote":"Grounds the Free Energy Principle used to justify the predict–observe–update loop for cognitive uncertainty."}],"review_version":1}