{"id":"85ca77f9-d986-4aee-a467-ee04aef7e68b","arxiv_id":"2412.14646","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":15,"one_line_summary":"A soft-feedback Bayesian message rule speeds up collective decisions in a swarm of miniaturized vibration-sensing robots while preserving accuracy, and PSO-optimized parameters generalize better to difficult floor patterns than hand-tuned ones.","lead":"This paper tests a new information-sharing rule, called soft feedback, for a swarm of tiny vibration-sensing robots that must decide whether most tiles on a floor are vibrating. The rule makes the swarm reach a collective decision faster than two established strategies, without losing accuracy, in both simulation and real-robot trials.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Unmodeled packet loss and single-layout calibration may invalidate simulated robustness and extrapolation of P* to complex 10x10 environments; real-world evidence covers only one 5x5 pattern.","rationale":"Read in good faith, the paper has real strengths: a physics-based Webots simulation calibrated on five features with cosine similarities of 0.89-0.98 (Section 7.1), a noise-resistant PSO procedure, and real-robot validation across swarm sizes 5-10. The soft-feedback message rule is simple and clearly specified. However, the principal support for 'consistently outperforms' is the simulation study; the real experiments cover only one fixed 5x5 pattern at f=0.48 (Section 7.4). The simulation's ranking of u- versus u+ is reversed by real experiments because of packet loss (Section 7.4, Figures 14 and 15d), demonstrating that the simulator's communication model is not faithful enough to support strategy ranking or parameter transfer when loss is present. Since the 10x10 pattern robustness study (Figure 12) is purely simulation-based and loss-free, the extrapolation of P* and the robustness claim are not yet grounded in evidence. This matches the reader's weakest assumption, so agreement is 'agree.' A conditional verdict remains appropriate; the missing piece is either a packet-loss-inclusive simulation or a real experiment on a high-Moran pattern. Secondary issues such as the unreported decision threshold pc, absent code/data release, and under-specified concurrent implementation are real but less load-bearing than the communication-model gap.","tokens_in":20866,"tokens_out":8834,"duration_ms":74565,"concrete_test":"Augment the Webots simulator with a per-message packet-loss model calibrated to Figure 15d (roughly linear from 0% loss at 5 robots to 7.5% at 10 robots, applied identically to u-, u+, and us), then rerun the Section 7.3 robustness suite: 5 robots, f=0.48 and f=0.46, patterns diagonal, stripe, block diagonal, alternating, and random, keeping P* and the soft-feedback parameters fixed. If us remains faster than u- and u+ and P* accuracy on diagonal and stripe stays at the Figure 12 level, the concern is resolved; if the strategy ordering reverses or P* accuracy drops below the u- baseline, the robustness and consistency claims fail.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central comparative claim rests on simulations that omit the real system's dominant communication failure mode. Section 7.4 and Figure 15d report 0-7.5% packet loss in RF24 communication, and this loss reverses the relative performance of u+ versus u- in real experiments compared with simulation (Section 7.4, Figures 14 and 11). The simulator therefore does not faithfully reproduce communication effects, yet the robustness of P* on diagonal, stripe, and block-diagonal high-Moran patterns (Section 7.3, Figure 12) is demonstrated only in this loss-free simulator, after calibration on a single 5x5 floor at f=0.48 with five robots (Section 7.1). Since soft feedback works by belief-dependent message randomization, packet loss changes which messages arrive and thus changes the feedback dynamics; the fact that u+/u- ordering flips shows that ranking strategies in the simulator is not reliable. Consequently, the 'consistently outperforms' claim is not established for realistic communication or for complex environments, and P* may not transfer. Real-robot data support soft feedback only on one floor pattern, so the gap between simulation and deployment is the load-bearing weakness.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper addresses a binary collective-perception task in a swarm of miniaturized vibration-sensing robots operating on a tiled surface. The robots use Bayesian inference on a Beta distribution over the fill ratio, and the paper proposes a third information-sharing strategy, soft feedback (us), in which each broadcast message is a random mixture of the robot's current observation and its belief-weighted decision. The authors calibrate a Webots simulation to their real robots using stochastic models of motor misalignment, battery drain, and sensing noise, optimize common algorithm parameters with a noise-resistant PSO, tune the soft-feedback parameters with a grid search, and compare us with no-feedback (u-) and positive-feedback (u+) strategies in simulation (100 runs per condition) and in real experiments with swarm sizes 5-10. The central claims are that soft feedback reduces decision time without reducing accuracy and that the optimized parameter set P* remains robust across floor patterns with different Moran indices and fill ratios.","tokens_in":21179,"tokens_out":9566,"duration_ms":69882,"significance":"If the claims hold, the paper makes a useful contribution to swarm robotics: a simple, communication-friendly feedback rule that accelerates Bayesian collective perception, supported by a calibrated physics-based simulation and by real-robot experiments across six swarm sizes. The strengths are the explicit calibration procedure with quantified cosine-similarity scores, the use of 100-run simulations with reported error bars, the noise-resistant PSO with staged re-evaluation, and the honest discussion of limitations, including the unmodeled packet loss and the single-pattern real-world validation. The paper does not overclaim theoretical novelty: it is an empirical optimization and validation study. Its significance is mainly as a demonstration that belief-dependent message randomization can speed up collective decisions in a realistic miniaturized-robot platform, and as a benchmark for future calibration-aware optimizations.","major_comments":[{"comment":"The robustness claim for P* and the broad claim of consistent outperformance rest on a simulator that assumes reliable all-to-all communication, while the real system exhibits 0-7.5% packet loss that, by the authors' own report, reverses the relative decision-time ordering of u+ versus u- compared with simulation (Section 7.4 and Figure 14 versus Figure 11). Since soft feedback changes the message distribution (Eqs. 16-18), dropped messages alter the belief dynamics in a way the simulator does not capture; the Discussion concedes that the calibrated simulation did not account for network loss. The measured loss rates should be incorporated as a sensitivity analysis (for example, Bernoulli message dropping in simulation), or the robustness claims should be explicitly restricted to loss-free conditions.","section":"Section 7.3, Figure 12; Section 7.4, Figure 15d; Section 8"},{"comment":"The decision threshold pc is an input to the algorithm and a central control of the speed-accuracy trade-off, but its value is never reported and it is absent from the PSO bounds and from the parameter set P*. This prevents reproduction of the results and comparison with the baseline algorithm of Ebert et al. (2020). Please report the value used in all experiments and simulations and provide a sensitivity analysis, or justify fixing it to a conventional value.","section":"Section 4, Eq. (13); Table 1"},{"comment":"The complex-environment simulations change the tile grid from 5x5 (20 cm tiles) to a 10x10 grid (10 cm tiles, assuming that is the intended meaning of the '10 x 10cm grid') while keeping the optimized parameters P*, whose sampling interval tau = 3778 ms was described as approximately one 20-cm-tile travel distance. Because robot speed and collision-avoidance thresholds are unchanged, the same parameter values imply different tile-relative motion and different spatial correlation of samples; this confounds the comparison between P* and P0 in Figure 12. The authors should either rescale the motion and sampling parameters to preserve tile-relative behavior, or rerun the comparison on environments that do not change the calibration scale.","section":"Section 7.3"},{"comment":"Real-robot evidence covers only one 5x5 floor pattern at f=0.48, and the calibration metrics are about motion and sensing statistics (state times, sample distributions, inter-sample distances), not about predicted decision times or accuracies. Therefore the statement that experimental findings are in line with simulated findings is not quantitatively supported for the decision variables themselves. Please include a matched simulation-versus-experiment comparison of decision time and accuracy for at least one common condition, or explicitly relabel the real experiments as a qualitative demonstration rather than a quantitative validation.","section":"Section 7.1 and Section 7.4"}],"minor_comments":[{"comment":"The phrase 'the environment is reduced to a 10 x 10cm grid' is ambiguous; please clarify whether this is a 10x10 grid with 10 cm tiles or a 10 cm by 10 cm surface.","section":"Section 7.3"},{"comment":"The sentence 'The probability of FP and FN are roughly equal in simulation and experiments, measuring 13%' should state whether 13% is the false-positive rate, the false-negative rate, or their average. It would also be clearer to use 'rate' rather than 'probability' here.","section":"Section 7.1"},{"comment":"The running RMS computation updates n before the summation, so with n=0 initially the first filtered sample is handled inconsistently; please fix the off-by-one in the index or initialize n=1.","section":"Algorithm 2"},{"comment":"The symbol Gamma is used for the variance of the Beta distribution in Eq. (17) and for the Gamma-distribution parameters in Eq. (20); using distinct notation, such as Var(Beta), would avoid confusion.","section":"Section 4.1, Eqs. (17) and (20)"},{"comment":"Because the central claim is 'consistently outperforms' and 'without compromising accuracy', the paper would benefit from reporting pairwise confidence intervals or significance tests on the decision-time and accuracy differences rather than only means and error bars.","section":"Figures 10, 11, 14, and 15"},{"comment":"The text contains a typo, 'an binary inspection problem', which should read 'a binary inspection problem'.","section":"Section 1"}],"recommendation":"major_revision","confidential_remarks":"The paper is a substantial extension of the authors' own ANTS 2024 conference paper, and this origin is disclosed. The main concern for me is that the robustness claim for P* is made in a simulator that omits the one communication failure mode that the real system exhibits and that demonstrably changes the ordering of the baseline strategies; this is fixable with a sensitivity analysis or by narrowing the claims. The missing decision-threshold value and the tile-size confound in the complex-environment simulations are also fixable. The central soft-feedback result is supported by direct real-robot evidence at six swarm sizes, so I do not see grounds for rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know this is a solid empirical paper, not a breakthrough. The soft feedback mechanism itself appeared in their own 2024 ANTS paper; what this preprint adds is a two-stage optimization pipeline, a more thoroughly calibrated Webots model, and a broader robustness study. The main claim—soft feedback cuts decision time without hurting accuracy—is supported by 100-run simulations and real experiments across six swarm sizes.\n\nThe calibration work is unusually thorough. They match state times, vibration distributions, sample spatial distributions, and travel distances with cosine similarities around 0.9. The PSO plus grid search is sensible, and they quantify the sim-to-real gap honestly, including the packet loss that they failed to model. That honesty matters.\n\nThe big soft spot: the simulator does not model network packet loss, and real experiments show 0–7.5% loss that flips the relative ordering of u+ vs u-. The stress-test worry that this invalidates all strategy rankings in simulation is too strong—the real experiments still show soft feedback faster than both alternatives. But it does undermine the robustness claim for P* on high-Moran 10x10 patterns, which is simulation-only. Calibration was done on one 5x5 layout at f=0.48 with five robots; extrapolating to 10x10 and other swarm sizes is a leap. Also, they never report the decision threshold pc, and the concurrent implementation of three strategies in real experiments is underspecified. These are fixable. No code or data release either.\n\nWho gets value: groups working on collective perception with real robot swarms, and people building sim-to-real pipelines for swarm optimization. It deserves a serious referee. The empirical effort is substantial, and the main flaw is addressable by adding packet loss to the simulator or qualifying the robustness claims. Rejecting would be wrong; accepting with major revisions is the right call.","headline":"A credible empirical extension of the authors' own earlier work; the central timing claim holds in real robots, but the simulation-only robustness claims are weaker than the paper implies.","tokens_in":21705,"tokens_out":2208,"would_cite":true,"duration_ms":17239,"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":"Soft feedback lets Bayesian robot swarms reach majority decisions faster while preserving accuracy.","keywords":["collective perception","swarm robotics","Bayesian decision-making","vibration sensing","soft feedback","particle swarm optimization","calibrated simulation","infrastructure inspection"],"falsifier":"A decisive test is to run the calibrated simulator for swarm size 10 on a high-Moran floor such as the stripe pattern (E_MI=0.88) with packet loss drawn from the measured 0-7.5% range, and compare soft-feedback decision time and accuracy against real experiments on the same layout; if soft feedback no longer beats u- and u+, or if P* accuracy drops below the experimental values, the claimed robustness fails.","tokens_in":20641,"feed_emoji":"🤖","tokens_out":6458,"duration_ms":42735,"temperature":0.7,"pith_summary":"This paper addresses whether a swarm of tiny vibration-sensing robots can collectively decide, faster than existing strategies, which type of tile covers most of a surface. The authors extend a Bayesian collective-perception algorithm by adding 'soft feedback,' a rule in which each broadcast mixes the robot's current observation with its evolving belief. They claim this rule consistently shortens decision time compared with the two established sharing strategies, no feedback and positive feedback, without sacrificing accuracy, in both calibrated simulation and real experiments. If true, it makes all-to-all Bayesian consensus a cheaper, more communication-friendly mechanism for mobile-sensor inspection.","feed_headline":"Soft feedback cuts swarm decision time by 17 percent.","feed_subtitle":"The new Bayesian sharing rule keeps accuracy intact while speeding up decisions across robot swarms in simulation and on real hardware.","key_machinery":"The load-bearing object is the Beta posterior each robot uses to model the surface fill ratio f, updated with local samples and broadcast messages. The new rule is the soft-feedback message sampler: m ~ Bernoulli(delta(1-p)+(1-delta)O) with delta = $e^{{-eta Gamma}}$|0.5-p|^kappa, where Gamma is the Beta variance and p is the robot's current belief that the surface is mostly non-vibrating. As confidence grows, the variance Gamma falls and delta rises, so messages gradually shift from raw observations toward belief-weighted decisions, pushing the swarm toward consensus without a hard irreversible commitment. A two-stage optimization supplies the parameter values: a noise-resistant PSO optimizes the common exploration and decision parameters to give P*, and a grid search selects soft-feedback coefficients eta=1500 and kappa=2. The calibrated physics-based simulator is what connects these parameter values to real robot behavior.","core_discovery":"The central claim is that soft feedback (u_s) outperforms the established no-feedback (u_-) and positive-feedback (u_+) strategies in decision time while preserving accuracy, across fill ratios from 0.44 to 0.56, swarm sizes 5 to 10, and environments with Moran index from -0.96 to 0.88. The paper also claims that the PSO-optimized parameter set P* = [7860, 10725, 3778, 55, 381] keeps accuracy stable across these environments, whereas the empirical parameter set P0 loses significant accuracy on clustered and structured floors such as diagonal, stripe, and block diagonal patterns. A 17% decision-time reduction is reported at the hardest fill ratio, and real experiments with up to 10 robots and packet loss up to 7.5% still show soft feedback as the fastest strategy. The intended consequence is that a simple randomized message rule can replace irreversible commitment as the consensus-driving mechanism.","pith_inferences":["If soft feedback's edge comes from replacing irreversible commitment with gradual belief-weighted mixing, then the same rule should accelerate other consensus algorithms that currently rely on positive feedback, such as voter models or majority-based quorum sensing, in all-to-all communication regimes.","A direct testable extension is to add packet-loss noise to the calibrated simulator; the paper's real data suggest this flips the u+ versus u- ordering and would let simulation match experiment.","Because P* was optimized for one fill ratio (0.48), re-optimizing for each swarm size or for f near 0.5 should close remaining accuracy gaps, an option the paper notes was computationally infeasible.","Soft feedback also functions as a privacy-preserving consensus mechanism: robots broadcast randomized messages that only statistically encode their beliefs, which could matter for distributed sensor networks with limited bandwidth."],"forward_implications":["Soft feedback reduces decision time by roughly 17% at the hardest fill ratio (f=0.48) compared with no-feedback, without lowering accuracy.","Optimized parameters P* keep accuracy stable across Moran indices from -0.96 to 0.88 and fill ratios 0.46-0.48, while empirical parameters P0 fail on clustered floors.","In real experiments with packet loss 0-7.5%, soft feedback remains the fastest strategy at all swarm sizes from 5 to 10, unlike positive feedback, whose advantage over no-feedback grows with packet loss.","Larger swarm sizes shorten decision time but do not improve accuracy, because collision avoidance increases the spatial correlation of samples.","The two-stage optimization approach—PSO for common parameters, grid search for soft-feedback gains—carries over to other Bayesian collective-decision settings."],"supporting_citations":[{"why":"Supplies the Bayesian Beta-distribution algorithm and the u- and u+ feedback strategies that this work extends.","marker":"Ebert et al. (2020)"},{"why":"Earlier conference version that introduced soft feedback and the initial robot and simulation setup, which this paper extends.","marker":"Siemensma et al. (2024)"},{"why":"Provides the Moran Index and entropy measures used to quantify environmental difficulty.","marker":"Bartashevich and Mostaghim (2019a)"},{"why":"Introduces the Rovable miniaturized robot platform used in the real experiments.","marker":"Dementyev et al. (2016)"},{"why":"Provides the linearly decreasing inertia weight for the PSO optimization used in the first optimization stage.","marker":"Shi and Eberhart (1998)"},{"why":"Develops the simulation framework that the paper's calibrated physics-based simulator builds upon.","marker":"Chiu, Nagpal, and Haghighat (2024)"}],"fun_headline_variants":["Soft feedback cuts swarm decision time by 17%","Bayesian swarm chooses faster with soft feedback","Randomized sharing rule speeds up robot swarm votes","Soft feedback outperforms commitment in swarm decisions","Mini robots decide faster using softer feedback"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a simulation calibrated with five robots on one 5x5 layout at fill ratio 0.48 remains faithful enough to predict performance for swarms of 5-10 robots and for 10x10 environments with high spatial correlation, even though the simulation does not model network packet loss.","fun_headline_variants_meta":{"raw":{"variants":["Soft feedback cuts swarm decision time by 17%","Bayesian swarm chooses faster with soft feedback","Randomized sharing rule speeds up robot swarm votes","Soft feedback outperforms commitment in swarm decisions","Mini robots decide faster using softer feedback"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000176,"raw_usage":{"total_tokens":1308,"prompt_tokens":984,"completion_tokens":324,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":600,"completion_tokens_details":{"reasoning_tokens":256}},"tokens_in":600,"tokens_out":324,"duration_ms":2766,"temperature":1.0,"reasoning_tokens":256,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T12:03:27.174608+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A decisive test is to run the calibrated simulator for swarm size 10 on a high-Moran floor such as the stripe pattern (E_MI=0.88) with packet loss drawn from the measured 0-7.5% range, and compare soft-feedback decision time and accuracy against real experiments on the same layout; if soft feedback no longer beats u- and u+, or if P* accuracy drops below the experimental values, the claimed robustness fails.","supporting_citations":[{"cited_title":", Gauci, M","cited_arxiv_id":null,"evidence_quote":"Supplies the Bayesian Beta-distribution algorithm and the u- and u+ feedback strategies that this work extends."},{"cited_title":", Chiu, D","cited_arxiv_id":null,"evidence_quote":"Earlier conference version that introduced soft feedback and the initial robot and simulation setup, which this paper extends."},{"cited_title":", Kao, H.L.C","cited_arxiv_id":null,"evidence_quote":"Introduces the Rovable miniaturized robot platform used in the real experiments."},{"cited_title":"\\ Eberhart, R.C","cited_arxiv_id":null,"evidence_quote":"Provides the linearly decreasing inertia weight for the PSO optimization used in the first optimization stage."},{"cited_title":", Nagpal, R","cited_arxiv_id":null,"evidence_quote":"Develops the simulation framework that the paper's calibrated physics-based simulator builds upon."}],"review_version":1}