{"id":"87717689-494d-4a27-9ba9-0d0089d0b8ab","arxiv_id":"2506.03847","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"Cooperative neuroevolution with CPPN-NEAT produced soft actuator morphologies with higher simulated bending and controller robustness than the AFPO baseline.","lead":"This paper evolves soft actuator shapes and controllers together using cooperative neuroevolution in a physics simulator. It reports better bending displacement and controller robustness than a prior AFPO baseline, with evidence limited to simulation and a single best morphology per method.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Single-morphology post hoc comparison: the claimed superiority of NF over AFPO rests on one fittest SAM per method, with AFPO's evolutionary history and search effort incomparable, so the central claim is underdetermined by the reported evidence.","rationale":"The reader identified the simulator-to-reality transfer as the weakest assumption, which is a genuine limitation but is largely a generalizability concern and is explicitly acknowledged by the authors as future work. My stress-test pass finds a more immediate, internal evidentiary weak point: the central comparative claim depends on a single post hoc selected morphology per method, with no matched search effort, no reported variance across AFPO runs, and a post hoc displacement threshold that excludes three of four collaboration strategies. This is a correctness-of-inference concern within the paper's own simulated domain; it does not depend on whether Voxelyze transfers to the real world. Because the evidence base for the superiority claim is thin and the artifacts (code, data, seeds) are not provided to allow re-analysis, the appropriate verdict remains CONDITIONAL—the claim is plausible but not established at the stated strength. I do not recommend REJECT because the experimental setup is internally coherent and the reported differences are statistically significant in the specific comparison performed; the issue is that the specific comparison performed is not sufficient to support the general conclusion. The reader's weakest_assumption partially overlaps with my concern (both point to external validity), but I emphasize a load-bearing internal comparison flaw, so agreement is 'partial' rather than 'agree.'","tokens_in":5239,"tokens_out":1709,"duration_ms":14947,"concrete_test":"Re-run the comparison with matched search effort and multiple seeds: fix the same total number of fitness evaluations for AFPO and for each coevolution strategy (e.g., 100,000 evaluations), run at least 10 independent evolutionary runs per method, and report the distribution of best-morphology displacement for each method rather than a single fittest SAM. If AFPO's best-morphology distribution overlaps or exceeds the coevolution distribution under matched evaluations, the paper's central claim fails. Additionally, report the pre-threshold performance of NW, NFW, and NR with the 0.1 threshold applied and removed, to confirm the superiority of NF is not an artifact of post hoc selection.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that cooperative neuro coevolution with NF is more suitable than AFPO for designing soft actuator morphologies (SAMs), based on two experiments: (i) identifying the best collaboration strategy under NF, and (ii) comparing the single best NF morphology against the single best AFPO morphology under 1000 random controllers. The load-bearing weakness is that the AFPO baseline is not an equivalent or well-matched control. First, the AFPO comparison uses only one morphology per algorithm: the fittest SAM from an unspecified number of AFPO runs (the paper cites reference [3] without reporting the number of runs, variance, or the evolutionary history). Second, the comparison does not control for search effort or population size: the cooperative scheme uses two populations of 25 individuals (50 total evaluations per round-robin generation), while AFPO uses 50 individuals in a single population; if the coevolution run is counted in terms of total fitness evaluations per generation, the NF approach may receive more or differently structured search effort. Third, the threshold of 0.1 displacement is applied post hoc to exclude NW, NFW, and NR from further analysis; the paper reports no independent or a priori justification for this threshold, so the comparison is effectively 'best-of-four strategies versus AFPO,' which is a selection-after-the-fact design. Fourth, no statistical comparison of the morphologies themselves is possible because the comparison is n=1 per algorithm. Finally, the paper defers friction and viscosity to future work, so the simulator-to-physical transfer is untested. These issues do not make the result false, but they make the stated conclusion that cooperative coevolution is 'more suitable' than AFPO much stronger than the evidence supports.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a cooperative neuroevolution approach, based on CPPN-NEAT, to simultaneously evolve the morphology of a soft actuator and its controller, with Voxelyze as the physics-based fitness evaluator. Four collaboration strategies are compared (n fittest, n worst, n fittest and worst, and n random individuals vs all), and the n fittest strategy is reported as best. The fittest morphology from this approach is then compared against a single AFPO baseline morphology from the authors' earlier work, under 1000 randomly generated phase-offset controllers, using upward bending displacement and robustness as metrics. The paper concludes that cooperative neuro coevolution with the n fittest collaboration strategy is more suitable for designing soft actuator morphologies than AFPO.","tokens_in":5475,"tokens_out":5917,"duration_ms":55918,"significance":"If the comparison were properly controlled, the result would be of practical interest for biohybrid soft actuator design, since morphology and controller search are coupled and the evaluation uses an independent physics simulator. The collaboration-strategy experiment uses 10 independent runs, and the robustness evaluation under 1000 controller variations is a concrete, falsifiable test. The central weakness is that the headline claim of superiority over AFPO rests on a single pair of morphologies, with unmatched search effort and a post hoc selection threshold, so the claim is not yet established by the reported evidence.","major_comments":[{"comment":"The headline comparison is based on one fittest SAM per algorithm. The manuscript does not report how many independent AFPO runs were performed in reference [3] or the variance of the AFPO best morphology, and the coevolution side selects one fittest SAM from the NF configuration. With n=1 per method, the observed difference in mean displacement cannot be attributed to the algorithm rather than to run-to-run chance. Please report distributions over multiple independent runs for both methods, and compare those distributions with an appropriate statistic that treats the run as the unit of replication.","section":"Section 3.2, Figure 2"},{"comment":"Search effort is not matched. The coevolutionary scheme uses two populations of 25 individuals evolved in round-robin fashion, while AFPO uses a single population of 50 individuals. Because the number of Voxelyze evaluations per generation differs between these designs, the comparison in Section 3.2 conflates algorithmic efficacy with total evaluation budget. Please report the number of fitness evaluations per generation for each method and run both under the same total evaluation budget.","section":"Section 2.3"},{"comment":"The displacement threshold of 0.1 is introduced only after observing that NW, NFW, and NR did not reach it, and no independent or a priori justification is given. This makes the subsequent 'NF vs AFPO' comparison a post hoc best-of-four selection. Please either state and justify the threshold before the experiment, or report all four collaboration strategies in the final comparison with appropriate multiple-comparison control.","section":"Section 3.1"},{"comment":"Figure 1 shows only mean curves over the 10 runs, with no confidence intervals or per-run variability. The statistical summary reports only thresholded p-values from Dunn's test; exact p-values and the multiple-comparison correction (for example, Bonferroni or Benjamini-Hochberg) are not stated. Without these, the ranking n=2 > n=3,n=5 > n=1 > n=10 is not fully supported.","section":"Section 3.1, Figure 1"},{"comment":"The paired Wilcoxon and t-tests compare 1000 controller draws for the two selected morphologies. These tests establish that these particular morphologies differ under random controllers; they do not establish that cooperative coevolution reliably produces more robust morphologies than AFPO. The authors should treat the morphology as a random effect by comparing across independent evolutionary runs.","section":"Section 3.2"}],"minor_comments":[{"comment":"The terms 'n best' and 'n fittest' are used interchangeably; please settle on one term throughout.","section":"Abstract and Section 2.3"},{"comment":"In the conclusions, the list of collaboration strategies repeats 'n fittest and worst individuals vs all' for item (d); the fourth strategy should be 'n random individuals vs all'.","section":"Section 4"},{"comment":"The vertical axis of the violin plots is labeled only as 'displacement observed in the yz plane'; please state the physical units or clarify that Voxelyze returns arbitrary length units.","section":"Figure 2"},{"comment":"It is not specified how the continuous CPPN outputs are decoded into the binary presence of a voxel and the discrete material type; please provide the decoding rule for reproducibility.","section":"Equation (1)"},{"comment":"The distribution of the 1000 random phase offsets is not described; please specify whether they are drawn uniformly over the controller output range and over what interval.","section":"Section 3.2"},{"comment":"The abstract says the approach 'can produce' more suitable morphologies, while the conclusions say it 'is more suitable'; please align the strength of the claim with the level of evidence.","section":"Conclusions"}],"recommendation":"major_revision","confidential_remarks":"The paper leans heavily on the authors' earlier work [2,3], and the AFPO baseline is imported from [3]. The editor may wish to ask the authors to clarify the novel contribution beyond those papers and to confirm that the AFPO results were produced under conditions comparable to the coevolution runs in this manuscript."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: this is a real but narrow incremental result. The new bit is comparing four cooperative coevolution collaboration strategies under CPPN-NEAT for soft actuator morphology design and finding n-fittest-vs-all best, then beating an AFPO baseline from their own prior work in Voxelyze simulation. That comparison is the whole paper, and the evidence is thinner than the conclusion.\n\nWhat's good: the setup is clean and reproducible in principle. Voxelyze is a standard simulator, the CPPN encoding is sensible, the two-population round-robin scheme is clearly described, and the four strategies are tested across n values with 10 runs. The ranking of n=2 > n=3=n=5 > n=1 > n=10 under NF is a concrete, falsifiable claim. The robustness check under 1000 random controller phase offsets is a reasonable way to test controller-transfer. The paper also admits that friction and viscosity are deferred.\n\nWhere it gets soft: the headline claim that coevolution is 'more suitable than AFPO' rests on one fittest morphology per algorithm. That's n=1 for the comparison, and the AFPO morphology comes from a previous paper with no reported variance or number of runs. The 0.1 displacement threshold that eliminates NW, NFW, and NR is applied after seeing the results, and no a priori justification is given; that weakens the 'optimal configuration' conclusion. The statistics on the main comparison (Wilcoxon, paired t-test) are fine for the 1000-sample robustness test, but they don't fix the single-morphology design. And no code or data are released, so independent verification of the actual morphologies is not possible.\n\nNone of this makes the result false. The simulation is a legitimate fitness function, and the central ranking among collaboration strategies is probably robust. But the 'more suitable than AFPO' sentence is a two-algorithm, one-morphology comparison, and the paper should be read as a proof-of-concept, not a decisive head-to-head.\n\nSend it to peer review. A serious referee can ask for code/data, multiple AFPO runs, and pre-registered thresholds; the underlying experiment is worth engaging. I'd cite it if I worked on neuroevolution for soft actuators, but I'd cite it as evidence about collaboration strategy ranking, not as evidence that coevolution beats AFPO.","headline":"A clean but narrow in silico comparison of coevolution strategies; the AFPO head-to-head overclaims what n=1 can support.","tokens_in":6078,"tokens_out":1668,"would_cite":true,"duration_ms":15556,"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":"Coevolving a soft actuator's shape and controller beats evolving its shape alone.","keywords":["cooperative coevolution","neuroevolution","NEAT","compositional pattern-producing networks","soft actuator","Voxelyze","morphology design","drug delivery"],"falsifier":"Run the fittest cooperative-coevolution morphology and the fittest AFPO morphology through the same 1000 phase-offset scenarios in a Voxelyze model with friction and viscosity enabled, or in a physical soft-actuator prototype; the central claim fails if the coevolved morphology does not show significantly higher mean upward displacement and steadier performance.","tokens_in":5037,"feed_emoji":"🦾","tokens_out":13671,"duration_ms":112308,"temperature":0.7,"pith_summary":"This paper argues that the best way to automate the design of soft actuator morphologies for drug-delivery devices is to evolve the morphology and its controller together in two interacting populations, rather than optimising a single population with a multi-objective algorithm. Under a cooperative neuroevolution scheme, the paper tests four collaboration strategies and reports that only evaluating every individual against the $n$ fittest individuals of the other population reliably produces upward bending above a 0.1 displacement threshold, with $n=2$ performing best. The fittest morphology from this scheme bends further and stays more consistent across 1000 randomly generated controller phase offsets than the fittest morphology found by Age-Fitness Pareto Optimisation. If this result transfers from the Voxelyze physics simulator to physical biohybrid actuators, it gives designers an automated path to bendier, more robust soft medical devices.","feed_headline":"Coevolution beats single-population search for soft actuator shapes","feed_subtitle":"Evolving shape and controller together yields bendier, steadier soft actuators for drug delivery","key_machinery":"The central machinery is a two-population cooperative coevolution of Compositional Pattern-Producing Networks (CPPNs). Morphology CPPNs are queried as $\\mathrm{CPPN_{sam}}(x_i,y_i,z_i) \\to (\\nu_i, m_i)$, reporting voxel presence and material type across a $20 \\times 8 \\times 8$ canvas; controller CPPNs are queried as $\\mathrm{CPPN_{con}}(x_i,y_i,z_i,m_i) \\to \\phi_i$, reporting the phase offset of each active voxel's contraction, clamped to $[-2\\pi,2\\pi]$. Voxelyze acts as the fitness function, simulating the mechanical response of active and passive voxels and tracing the free end's displacement. The collaboration strategy determines which members of one population evaluate the other: the paper contrasts the $n$ fittest, $n$ worst, $n$ fittest and worst, and $n$ random individuals against all.","core_discovery":"The paper's central claim is that Neuroevolution of Augmented Topologies (NEAT) driving Compositional Pattern-Producing Networks (CPPNs) is more suitable for designing soft actuator morphologies when embedded in a cooperative coevolutionary scheme using the n fittest individuals vs all collaboration strategy than when run as the single-population Age-Fitness Pareto Optimisation baseline. In the authors' Voxelyze simulations, the fittest coevolved morphology shows higher maximum upward bending displacement and, across 1000 controller phase-offset scenarios, a higher and more concentrated displacement distribution than the AFPO morphology. The authors attribute the advantage to the mutual evolutionary pressure between the morphology and controller populations, and to CPPNs' ability to generate patterns such as the solid diagonal arrangement of active voxels in the winning morphology.","pith_inferences":["The paper compares one winning morphology per algorithm, so whether cooperative coevolution reliably dominates AFPO across many independent runs or across whole morphology populations is left untested; the advantage could belong to the fittest designs rather than to the search scheme.","Because fitness is the arithmetic mean over collaborators and controller outputs are clamped to $[-2\\pi,2\\pi]$, the reported ranking of collaboration sizes could shift under worst-case or rank-based fitness aggregation, a testable variant the paper does not run.","The fixed passive bioreactor enclosure and the omission of friction and viscosity mean the diagonal active-voxel morphology may be optimised for the simulator rather than for physical conditions; adding those effects, which the paper lists as future work, could change the ranking and the recommended shape.","A stronger robustness test would pit the evolved morphology against adversarially constructed controllers rather than random phase offsets, better mimicking a real environment where control signals actively distort."],"forward_implications":["The default configuration for this design task should be cooperative coevolution with the two fittest individuals of one population evaluating all individuals of the other, since $n=2$ ranked above $n=1,3,5,10$ in the reported comparisons.","Morphology and controller should be designed jointly, because the mutual selection pressure between the two populations is what produces bendier structures; separate evolution runs are expected to lag.","The winning morphology's solid diagonal arrangement of active voxels can be treated as a candidate design heuristic for upward bending, replacing human intuition about where to place active tissue.","Morphologies that score well under 1000 random controller phase offsets are expected to tolerate signal cross-talk in real operation, because the robustness metric directly tests a fixed morphology against many controllers."],"supporting_citations":[{"why":"supplies the Voxelyze physics engine used to simulate voxel mechanics and compute fitness","marker":"[7]"},{"why":"provides the NEAT algorithm that evolves both CPPN populations with topology augmentation","marker":"[13]"},{"why":"defines Compositional Pattern-Producing Networks, whose periodic functions generate morphological patterns","marker":"[11]"},{"why":"supplies the cooperative coevolution methodology and the round-robin population scheme","marker":"[8]"},{"why":"defines Age-Fitness Pareto Optimisation, the baseline algorithm the paper compares against","marker":"[9]"},{"why":"provides the AFPO-evolved soft actuator morphology and prior experimental setup used in the comparison","marker":"[3]"},{"why":"supports clamping controller outputs to [-2π, 2π] for full sinusoidal contractions","marker":"[1]"},{"why":"conceptualises the passive bioreactor enclosure that the simulated actuator includes","marker":"[14]"}],"fun_headline_variants":["Coevolving morphology and controller beats AFPO","Cooperative NEAT designs better soft actuators","Pairing shape and control evolution improves soft robots","Coevolution outperforms AFPO for soft drug delivery bots","Joint evolution of body and brain wins for soft actuators"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that Voxelyze's simulation of contracting active and passive voxels is a faithful proxy for how a real biohybrid soft actuator will bend; if friction, viscosity, or biological tissue behaviour reverse the ranking, the reported advantage of cooperative coevolution over AFPO will not hold in physical devices.","fun_headline_variants_meta":{"raw":{"variants":["Coevolving morphology and controller beats AFPO","Cooperative NEAT designs better soft actuators","Pairing shape and control evolution improves soft robots","Coevolution outperforms AFPO for soft drug delivery bots","Joint evolution of body and brain wins for soft actuators"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000303,"raw_usage":{"total_tokens":1714,"prompt_tokens":890,"completion_tokens":824,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":506,"completion_tokens_details":{"reasoning_tokens":752}},"tokens_in":506,"tokens_out":824,"duration_ms":8357,"temperature":1.0,"reasoning_tokens":752,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T10:54:08.529575+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the fittest cooperative-coevolution morphology and the fittest AFPO morphology through the same 1000 phase-offset scenarios in a Voxelyze model with friction and viscosity enabled, or in a physical soft-actuator prototype; the central claim fails if the coevolved morphology does not show significantly higher mean upward displacement and steadier performance.","supporting_citations":[{"cited_title":"Kriegman","cited_arxiv_id":null,"evidence_quote":"supplies the Voxelyze physics engine used to simulate voxel mechanics and compute fitness"},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"provides the NEAT algorithm that evolves both CPPN populations with topology augmentation"},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"defines Compositional Pattern-Producing Networks, whose periodic functions generate morphological patterns"},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies the cooperative coevolution methodology and the round-robin population scheme"},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"defines Age-Fitness Pareto Optimisation, the baseline algorithm the paper compares against"},{"cited_title":"Alcaraz-Herrera, M.-A","cited_arxiv_id":null,"evidence_quote":"provides the AFPO-evolved soft actuator morphology and prior experimental setup used in the comparison"},{"cited_title":"Alcaraz-Herrera, M","cited_arxiv_id":null,"evidence_quote":"supports clamping controller outputs to [-2π, 2π] for full sinusoidal contractions"},{"cited_title":"Tsompanas and I","cited_arxiv_id":null,"evidence_quote":"conceptualises the passive bioreactor enclosure that the simulated actuator includes"}],"review_version":1}