{"id":"ecd182e9-258b-48cf-bbcf-4247b72069e5","arxiv_id":"2608.10056","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":11,"one_line_summary":"A constrained reinforcement learning framework with separate proximity and safety cost limits, plus conformal prediction uncertainty margins, improves human-following robots in dense crowds.","lead":"A robot-following system splits 'stay close' and 'stay safe' into separate constraint limits instead of one blended reward, and adds uncertainty-aware safety margins around predicted pedestrian positions. In simulations and one real-robot test, it reports better following and collision trade-offs than baselines.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (5) assigns zero following cost for d_follow < d_personal, so the claimed behavioral meaning of delta_F (moderate distance, no personal-space intrusion) is not encoded; the central control claim rests on a cost that does not penalize crowding.","rationale":"I read the paper as trying to establish that cost thresholds with behavioral meanings provide explicit, predictable control over the proximity-safety trade-off, superior to reward-weight tuning. The reader's formal weakest assumption is the reliability of the ACI uncertainty bounds, but the paper's own rationale already identifies the Eq. (5) dead zone. That internal inconsistency is more load-bearing because it attacks the claimed mechanism: if the following cost does not penalize crowding, then delta_F cannot mean 'desired following distance' or 'avoid personal-space intrusion' as claimed. The ACI issue is important but pertains to whether the safety cost truly encodes collision risk; the Eq. (5) issue pertains to whether the primary demonstration of threshold control measures what it says. A corrected symmetric cost might preserve the empirical trends, so the appropriate verdict remains CONDITIONAL rather than REJECT, but as written the central claim is not supported for the following-distance dimension. I also note the self-reported best-of-tuning selection in Sec. IV.C and the absence of error bars, which the reader correctly flags.","tokens_in":18847,"tokens_out":5550,"duration_ms":53764,"concrete_test":"Modify Eq. (5) to a symmetric penalty, e.g., C_F = k1 * |d_follow - d_personal|, retrain the three Table II profiles (delta_F = 4.0/3.6/3.2 with the corresponding delta_H values) under otherwise identical settings, and recompute AFD and TLR. If the monotonic relationship between delta_F and AFD/TLR persists, the central claim survives with a corrected cost; if the trade-off curve changes or moves by more than seed noise, the reported 'direct behavioral control' was an artifact of the one-sided cost. Alternatively, before retraining, evaluate the policy at delta_F=3.2 and measure the fraction of timesteps with d_follow < 1.0 m: any non-negligible fraction would directly demonstrate the dead-zone exploitation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section III.D states that C_F 'encourages the robot to maintain a moderate following distance rather than either lagging behind or crowding the target' and imposes the equality constraint E[sum C_F] = delta_F. Equation (5), however, defines C_F = k1*(d_follow - d_personal) for d_follow > d_personal and 0 otherwise. Distances below d_personal are completely free of cost. A policy can therefore drive arbitrarily close to the target at zero following cost and still satisfy a positive delta_F by sometimes being far away; nothing in the objective penalizes personal-space intrusion. Thus delta_F does not directly determine a moderate following distance; at most it controls the average excess distance above 1.0 m, and the equality-constrained Lagrangian update has no signal to prevent dead-zone exploitation. The central claim of the paper, that each cost threshold carries direct behavioral meaning and provides explicit, predictable control over the proximity-safety trade-off, is unsupported for the following dimension as written. This is compounded by Section IV.C's admission that reported results are 'the best result among all tuned configurations' for both the proposed method and RL+ACI, which biases the headline comparisons; but the Eq. (5) issue is internal and strikes at the claimed mechanism itself.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a constrained reinforcement learning (CRL) framework for human-following robots in crowded pedestrian environments. The task is decomposed into a sparse task reward and three independent cost constraints—following distance, human safety, and obstacle safety—with thresholds δ_F, δ_H, and δ_O intended to have direct behavioral meaning. Prediction uncertainty of human motion is quantified via adaptive conformal inference (ACI) and integrated into both the observation space and the human-safety cost. The method is evaluated in an extended CrowdNav simulator with static obstacles, across in-distribution and out-of-distribution scenarios, and deployed on a real robot. The central claim is that cost-threshold tuning provides explicit and predictable control over the proximity-safety trade-off, in contrast to implicit reward-weight tuning.","tokens_in":19096,"tokens_out":10156,"duration_ms":91013,"significance":"If the threshold-control claim held, the framework would offer a practical and interpretable alternative to dense reward shaping in human-following navigation, and the integration of ACI-based uncertainty into CRL costs is a sensible and potentially useful design. The paper contains substantial experimental work: multiple classical and learning baselines, four OOD scenarios, ablations, and a real-robot deployment. However, the central mechanism for the following-distance constraint is flawed as written (Eq. (5) has a dead zone below d_personal), and the experimental reporting is weakened by the selection of the 'best result among all tuned configurations' and the absence of variance statistics despite five seeds. The uncertainty-integration contribution also overlaps heavily with the authors' prior work [20,22]. With revisions, the framework could become a solid contribution, but the current manuscript does not substantiate its central interpretability claim.","major_comments":[{"comment":"The following cost is defined as C_F = k1(d_follow - d_personal) for d_follow > d_personal and 0 otherwise, so distances below d_personal incur no penalty. The equality constraint E[sum_t C_F] = δ_F can therefore be satisfied by a policy that spends some time arbitrarily close to the target and some time far away, with no cost signal preventing personal-space intrusion. The claim in Section IV.F that 'Tightening δ_F produces closer following' is not supported by this cost definition; at most δ_F bounds the average excess distance above d_personal. The manuscript should either modify C_F to penalize both sides of the desired distance (or add a separate personal-space constraint) and rerun the experiments, or substantially weaken the interpretability claims for δ_F.","section":"III-D, Eq. (5)"},{"comment":"Section IV.C states that for both RL+ACI and the proposed method the authors 'report the best result among all tuned configurations,' and Section IV.D says 1250 test samples across 5 seeds are used. Tables I–III report only point estimates with no standard deviations or error bars. As a result, the claimed improvements (e.g., SR 78.08% vs 71.60% for RL+ACI in Table I, and the monotonic trends in Table II) cannot be separated from tuning luck and seed variance. Please report mean ± std (or confidence intervals) across seeds for all metrics, disclose the full set of tuned configurations, and state how 'best' was chosen; otherwise the experimental comparisons are not statistically interpretable.","section":"IV.C and IV.D"},{"comment":"The human-safety cost C_H is built on ACI uncertainty bounds δ̂_{h,k}(t), and the paper claims improved safety in unpredictable OOD scenarios. However, no empirical coverage statistics for the ACI bounds are reported for the OOD conditions (rushing humans, SF model, groups, corridor), where the exchangeability assumptions underlying conformal prediction are violated. Without coverage or calibration data, the interpretation of δ_H as a safety threshold with direct behavioral meaning is not verified. Please report per-scenario empirical coverage, average error-bound ratios, and, if possible, conditional coverage as a function of horizon and crowd density.","section":"III-D and IV.G"},{"comment":"The uncertainty-aware cost formulation is listed as a main contribution, but Refs [20] and [22], both from the same research group, already present ACI-based uncertainty bounds integrated into constrained RL costs for crowd navigation. The manuscript should explicitly differentiate the present contribution from these works (e.g., the human-following setting, the interaction between the following constraint and the uncertainty-aware costs, or any architectural changes). As written, the novelty of the uncertainty-integration component is overstated, and the relation to prior self-authored work should be clarified in both the related work and the contributions list.","section":"III-C,D and Contributions"}],"minor_comments":[{"comment":"The indicator notation in Eq. (2) is easy to misread; please define the indicator explicitly (e.g., I[·]) and clarify the dependence on h and k.","section":"Eq. (2)"},{"comment":"The column grouping under 'CR' should be labeled explicitly (e.g., 'CR (Overall / Human / Obstacle)') to avoid ambiguity.","section":"Tables I and III"},{"comment":"Section IV.H reports 10 real-robot trajectories with 7 successes but does not state the criteria for success, the number of moving pedestrians, or the specific perception failures; a short table or bullet list would help substantiate the feasibility claim.","section":"IV.H"},{"comment":"The paper does not mention whether code and simulation environments will be released; for reproducibility, please add a statement or public repository link on the project page.","section":"Section IV.D"},{"comment":"In Eq. (12), λ_F is updated without a non-negativity clamp, unlike Eq. (11); please clarify whether negative multipliers are allowed and how the equality constraint is enforced in practice.","section":"Eq. (12)"}],"recommendation":"major_revision","confidential_remarks":"The overlap with Refs [20] and [22] is significant; the authors should be asked to state explicitly what is new. Also, the reporting of 'best result among all tuned configurations' is concerning; I would like the editor to require full tuning details and per-seed statistics before considering acceptance. The paper's scope fits the journal, but the central following-cost issue is a correctness problem that must be resolved."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick verdict: this is a solid systems paper with a real empirical payload, but the headline claim—thresholds with direct behavioral meaning—doesn't survive contact with its own following cost. Equation (5) charges zero cost when the robot is closer than d_personal = 1.0 m, so a policy can sit in the target's personal space for free, and the equality constraint on delta_F only controls average excess distance above that line. The abstract and Section IV.F claim explicit, predictable control over the proximity-safety trade-off via delta_F; that's only partially true. The human and obstacle collision constraints are more defensible, and the OOD results are genuinely useful.\n\nWhat is actually new: applying constrained RL with adaptive conformal inference to human following in crowds, with static-obstacle simulation and OOD evaluation. The Transformer+CNN policy is not new by itself, but the combination of three cost critics with ACI-derived uncertainty bounds is a working system. The comparison suite against SG-HA*, SG-ORCA, SG-MPC, OGM-HEIGHT, RL, and RL+ACI is extensive. The real-robot deployment with 7/10 successes and honest attribution of failures to perception is a plus.\n\nSoft spots: Section IV.C reports 'the best result among all tuned configurations' for both RL+ACI and the proposed method. That selection bias, combined with no error bars despite five seeds, makes the headline margins unreliable. The RL+ACI reward-weight tuning comparison in Table II rests on the same best-config reporting, so the claim that cost thresholds are more predictable than reward weights is not yet proven. The following-cost dead zone is not minor: it strikes at the paper's central mechanism. Unless the cost is changed to penalize deviations from a desired following range, or the claim is softened to 'delta_F controls the upper tail,' the interpretability story is overstated.\n\nSelf-citation note: the uncertainty-in-cost idea is largely from refs [20] and [22] by the same group. They are cited, but the incremental novelty here is modest. That doesn't bother me much; the system-level integration is still new.\n\nWho should read: people working on crowd navigation, safe RL, or human-robot following. The paper deserves a serious referee, not a desk reject, because the system works and the direction is valuable. But I'd send it back for a fix to Eq. (5), a variance-aware evaluation with fixed policies, and release of code and hyperparameters before the central claim can be accepted.","headline":"A solid systems paper with a real empirical payload, but the following-cost dead zone in Eq. (5) breaks the paper's central claim that thresholds provide direct behavioral control.","tokens_in":19742,"tokens_out":3837,"would_cite":true,"duration_ms":34924,"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":"Decomposing robot following into a sparse reward and three cost constraints gives explicit control over the proximity-safety trade-off.","keywords":["human-following robot","constrained reinforcement learning","cost constraints","proximity-safety trade-off","adaptive conformal inference","pedestrian crowds","PPO-Lagrangian","uncertainty-aware navigation"],"falsifier":"Run the trained policy in a crowd where pedestrians make abrupt 90-degree direction changes and measure, over all prediction horizons, how often the true prediction error exceeds the adaptive conformal bound; if the empirical coverage falls well below the nominal $1-\\alpha$ level, the safety cost has stopped encoding collision risk and the claimed safety behavior is not guaranteed.","tokens_in":18550,"feed_emoji":"🤖","tokens_out":12495,"duration_ms":100163,"temperature":0.7,"pith_summary":"This paper argues that the conflict between staying close to a target human and avoiding collisions in crowds is best managed by separating the task into a sparse reward for following and three independent cost constraints for following distance, human safety, and obstacle safety. Instead of tuning reward weights, the designer sets cost thresholds—a desired following distance, a tolerable intrusion level, and a safe obstacle clearance—that map directly onto behavior. The paper shows in simulation that shifting these thresholds moves the proximity-safety balance in the intended direction, while doubling reward weights can leave behavior unchanged or make it worse. It also feeds online prediction-uncertainty estimates into the human-safety cost, so the robot keeps a wider buffer when pedestrian motion is hard to predict. If correct, the framework gives robot designers a verifiable way to set how aggressively a follower robot should behave in dense, unfamiliar crowds.","feed_headline":"Three cost limits tune a follower robot's proximity-safety balance","feed_subtitle":"Decomposing the task into a sparse reward plus three constraints makes the trade-off explicit and predictable.","key_machinery":"The load-bearing object is the constraint decomposition itself: the CMDP in Eq. (3) with reward $R_t$ and costs $C^F_t$, $C^H_t$, $C^O_t$ subject to thresholds $\\delta_F$, $\\delta_H$, $\\delta_O$, optimized via PPO-Lagrangian with four critics and a combined advantage. The supporting mechanism is the adaptive conformal inference bound $\\hat\\delta_{h,k}(t)$ from Eq. (2), which is inserted both into the observation tokens and into the human-safety cost, where it expands the radius of safety regions around predicted pedestrian positions by exactly the current error bound. The attention-based Transformer policy fuses robot, target, obstacle, and human tokens so spatial constraints and social interactions are reasoned over jointly; at inference only the actor remains, so the decomposition costs nothing extra at deployment.","core_discovery":"The central claim is that in a constrained Markov decision process, a sparse task reward plus three independent cost constraints with behaviorally meaningful thresholds yields a proximity-safety balance that is explicit, tunable, and more predictable than the balance produced by a dense reward with tuned weights. The following cost uses an equality constraint to hold a desired distance $d_{personal}$, while human and obstacle collision costs use inequality constraints with limits $\\delta_H$ and $\\delta_O$; each cost is evaluated by its own critic and the policy is trained with PPO-Lagrangian. Prediction uncertainty from an adaptive conformal inference module, written $\\hat\\delta_{h,k}(t)$ for human $h$ at horizon $k$, expands the safety region around each predicted position when prediction errors grow. Direct comparisons show that changing $\\delta_F$ and $\\delta_H$ shifts average following distance and collision rates in the expected directions, whereas doubling reward weights fails to produce the intended effect and can degrade both following and safety.","pith_inferences":["Editorial inference: the same decomposition could be applied to other navigation or mobile-manipulation tasks with competing objectives, such as social navigation with personal-space preferences, replacing weight ratios with thresholds.","Editorial inference: because the ACI module only needs observed prediction errors, a more accurate pedestrian predictor should shrink the uncertainty buffers and permit closer following at the same safety level, which is a testable prediction.","Editorial inference: the equality constraint on the following cost may make training infeasible when $\\delta_F$ is set below what the robot's dynamics allow; the paper reports tuned values, but mapping the feasible range of $\\delta_F$ would clarify how far the tuning promise extends.","Editorial inference: the real-robot failures were attributed to upstream perception rather than the policy, so a testable extension is to feed detection confidence into the same uncertainty-aware cost structure."],"forward_implications":["A designer can specify the desired behavior directly by setting $\\delta_F$, $\\delta_H$, and $\\delta_O$: lower $\\delta_F$ gives closer following, lower $\\delta_H$ gives safer human interaction, and each moves the other in the expected direction.","Because thresholds map predictably to metrics, a designer can pre-select the intended behavior in simulation and then deploy the same thresholds, rather than hoping tuned weights generalize.","The uncertainty-aware cost extends to out-of-distribution crowds such as bidirectional corridors, rushing pedestrians, social-force groups, and moving groups, where the method keeps collision rates below non-uncertainty baselines.","At inference the actor network alone is used, so the explicit trade-off control comes with no additional runtime cost compared with single-critic RL.","Reward-weight tuning can fail in a counterintuitive way, as doubling the following weight left following distance unchanged and raised both target-loss and collision rates; cost thresholds avoid this failure mode."],"supporting_citations":[{"why":"Supplies the PPO-Lagrangian update rule that enforces the three cost constraints.","marker":"[9]"},{"why":"Provides the adaptive conformal inference bounds that expand safety regions around predicted pedestrian positions.","marker":"[6]"},{"why":"Grounds the use of adaptive conformal uncertainty in crowd navigation, motivating the uncertainty-aware cost design.","marker":"[20]"},{"why":"Provides the crowd navigation simulator used for training and in-distribution testing, extended with static obstacles.","marker":"[10]"},{"why":"Defines the constrained Markov decision process formulation that the task decomposition is built on.","marker":"[25]"},{"why":"Provides the PPO clipping objective used to update the actor with the combined advantage signal.","marker":"[30]"},{"why":"Supplies generalized advantage estimation, used to compute the separate reward and cost advantages.","marker":"[27]"},{"why":"Is the RL navigation baseline against which the method is compared, retrained with occupancy-grid inputs.","marker":"[29]"},{"why":"Is a classic velocity-based collision-avoidance baseline used in the comparisons.","marker":"[31]"}],"fun_headline_variants":["Sparse reward plus three constraints: explicit trade-off for robot followers","Robot follower balance: constraints beat reward weights for clarity","Decomposing human-following: sparse reward, tunable cost constraints","Predictive safety: uncertainty-aware cost limits improve crowd following","Three cost thresholds make proximity-safety trade-off explicit"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The human-safety cost carries collision risk only if the uncertainty estimates around predicted pedestrian positions truly bound the real prediction errors under the crowd behaviors encountered, including sudden turns and rushing pedestrians.","fun_headline_variants_meta":{"raw":{"variants":["Sparse reward plus three constraints: explicit trade-off for robot followers","Robot follower balance: constraints beat reward weights for clarity","Decomposing human-following: sparse reward, tunable cost constraints","Predictive safety: uncertainty-aware cost limits improve crowd following","Three cost thresholds make proximity-safety trade-off explicit"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000632,"raw_usage":{"total_tokens":2924,"prompt_tokens":956,"completion_tokens":1968,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":572,"completion_tokens_details":{"reasoning_tokens":1883}},"tokens_in":572,"tokens_out":1968,"duration_ms":11422,"temperature":1.0,"reasoning_tokens":1883,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T04:14:27.368293+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the trained policy in a crowd where pedestrians make abrupt 90-degree direction changes and measure, over all prediction horizons, how often the true prediction error exceeds the adaptive conformal bound; if the empirical coverage falls well below the nominal $1-\\alpha$ level, the safety cost has stopped encoding collision risk and the claimed safety behavior is not guaranteed.","supporting_citations":[{"cited_title":"Towards generalizable safety in crowd navigation via conformal uncertainty handling,","cited_arxiv_id":null,"evidence_quote":"Grounds the use of adaptive conformal uncertainty in crowd navigation, motivating the uncertainty-aware cost design."},{"cited_title":"Crowd-robot interaction: Crowd-aware robot navigation with attention-based deep reinforce- ment learning,","cited_arxiv_id":null,"evidence_quote":"Provides the crowd navigation simulator used for training and in-distribution testing, extended with static obstacles."},{"cited_title":"High- dimensional continuous control using generalized advantage estima- tion,","cited_arxiv_id":null,"evidence_quote":"Supplies generalized advantage estimation, used to compute the separate reward and cost advantages."},{"cited_title":"Height: Heterogeneous interac- tion graph transformer for robot navigation in crowded and constrained environments,","cited_arxiv_id":null,"evidence_quote":"Is the RL navigation baseline against which the method is compared, retrained with occupancy-grid inputs."}],"review_version":1}