Pith. sign in

REVIEW 5 major objections 6 minor 16 references

Designing morphologies of soft medical devices using cooperative neuro coevolution

T0 review · 5 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Coevolving a soft actuator's shape and controller beats evolving its shape alone.

desk verdict A clean but narrow in silico comparison of coevolution strategies; the AFPO head-to-head overclaims what n=1 can support. read the letter →

arxiv 2506.03847 v1 pith:CEEHCCCH submitted 2025-06-04 cs.NE

Hugo Alcaraz-Herrera , Michail-Antisthenis Tsompanas , Igor Balaz , Andrew Adamatzky This is my paper
classification cs.NE
keywords cooperativecoevolutionneuroevolutionNEATcompositionalpattern-producingnetworkssoftactuatorVoxelyzemorphologydesigndrugdelivery
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

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.

What carries the argument

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.

What would settle it

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.

Watch

Extended reading notes

Core claim

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.

Load-bearing premise

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.

Editorial extensions

If this is right

  • 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.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • 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.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 6 minor

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.

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 (5)
  1. [Section 3.2, Figure 2] 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.
  2. [Section 2.3] 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.
  3. [Section 3.1] 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.
  4. [Section 3.1, Figure 1] 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.
  5. [Section 3.2] 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.
minor comments (6)
  1. [Abstract and Section 2.3] The terms 'n best' and 'n fittest' are used interchangeably; please settle on one term throughout.
  2. [Section 4] 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'.
  3. [Figure 2] 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.
  4. [Equation (1)] 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.
  5. [Section 3.2] 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.
  6. [Conclusions] 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.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central comparison is an empirical evaluation against an independently evolved baseline in a physics simulator, not a reduction of the conclusion to its inputs.

full rationale

The paper's derivation chain is empirical rather than definitional. SAM morphologies and controllers are encoded as CPPNs, evolved under cooperative coevolution, and evaluated by the Voxelyze physics engine, which acts as the fitness function. The central claim—that cooperative neuro coevolution with the n-fittest collaboration strategy outperforms AFPO—rests on comparing the fittest evolved SAMs under 1000 randomly generated controller phase offsets. The AFPO baseline morphology is taken from the authors' prior work [3], but it is an independently evolved result from a different algorithm, not a parameter fitted in this paper to force the comparison. No equation defines the reported displacement in terms of an input parameter, and the objective metric is the same as the optimization target, which is standard practice rather than circular reasoning. The self-citations ([1], [2], [3]) provide implementation details, experimental infrastructure, and the baseline result; they do not smuggle in a uniqueness theorem or an ansatz that defines the outcome. Concerns about single-morphology comparison, unmatched search effort, and a post hoc displacement threshold are experimental-design weaknesses, not circularity. Therefore, no circular step is present.

Assumptions & free parameters 1 free parameters · 3 assumptions · 0 invented entities

The central comparison rests on simulation fidelity, on metric choice, and on comparability of the AFPO baseline. No free parameters are fitted to data; the main hand-chosen threshold is the displacement cutoff of 0.1 used to drop three collaboration strategies. No invented entities are introduced.

free parameters (1)
  • Displacement threshold = 0.1
    In Section 3.1, collaboration strategies NW, NFW, and NR are excluded because they did not reach a displacement threshold of 0.1; the threshold is not justified and no sensitivity analysis is provided.
assumptions (3)
  • domain assumption Voxelyze produces a valid mechanical response for active and passive voxels.
    Section 2.1 uses Voxelyze as the fitness function; all morphological improvements are measured inside this simulator, and the paper does not compare against physical experiments.
  • domain assumption Upward bending displacement and robustness to random phase offsets are appropriate proxies for drug delivery performance.
    Sections 1 and 3.2 define device suitability through these metrics; no drug delivery mechanism, fluid environment, or biological tissue model is simulated.
  • domain assumption The AFPO morphology from reference [3] is a directly comparable baseline.
    Section 3.2 compares against the fittest morphology found by AFPO in earlier work instead of rerunning AFPO under identical conditions; run-to-run or setup differences could confound the comparison.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Designing morphologies of soft medical devices using cooperative neuro coevolution." pith.science (2026). https://pith.science/paper/CEEHCCCH

@misc{pith2026250603847,
  author       = {Pith},
  title        = {Pith review of: Designing morphologies of soft medical devices using cooperative neuro coevolution},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CEEHCCCH}},
  note         = {Machine review of arXiv:2506.03847}
}
read the original abstract

Soft robots have proven to outperform traditional robots in applications related to propagation in geometrically constrained environments. Designing these robots and their controllers is an intricate task, since their building materials exhibit non-linear properties. Human designs may be biased; hence, alternative designing processes should be considered. We present a cooperative neuro coevolution approach to designing the morphologies of soft actuators and their controllers for applications in drug delivery apparatus. Morphologies and controllers are encoded as compositional pattern-producing networks evolved by Neuroevolution of Augmented Topologies (NEAT) and in cooperative coevolution methodology, taking into account different collaboration methods. Four collaboration methods are studied: n best individuals, n worst individuals, n best and worst individuals, and n random individuals. As a performance baseline, the results from the implementation of Age-Fitness Pareto Optimisation (AFPO) are considered. The metrics used are the maximum displacement in upward bending and the robustness of the devices in terms of applying to the same evolved morphology a diverse set of controllers. Results suggest that the cooperative neuro coevolution approach can produce more suitable morphologies for the intended devices than AFPO.

Figures

Figures reproduced from arXiv: 2506.03847 by the authors.

Figure 1
Figure 1. Mean performance observed of the fittest SAM using arithmetic mean under: (a) [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Displacement observed in the yz plane under 1000 phase offset scenarios of the fittest SAM found under AFPO (left); coevolution (right). 3.1 Optimal cooperative configuration The four collaboration strategies described in Section 2.3 are implemented in this experiment. For each collaboration strategy, individuals are evaluated using the arithmetic mean of the fitness of the corresponding collaborating individuals [… view at source ↗
Figure 3
Figure 3. Fittest SAM found by: (a) AFPO and (b) coevolution. Red voxels represent active tissue, whereas blue voxels [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

16 extracted references · 15 canonical work pages

  1. [3]

    Alcaraz-Herrera, M.-A

    H. Alcaraz-Herrera, M.-A. Tsompanas, I. Balaz, and A. Adamatzky. Using neuroevolution for designing soft medical devices.Biomimetic Intelligence and Robotics, 5(1), 2024

    work page 2024

  2. [1]

    Alcaraz-Herrera, M

    H. Alcaraz-Herrera, M. Tsompanas, I. Balaz, and A. Adamatzky. Control of biohybrid actuators using neuroevolu- tion. InProceedings of the 16th International Joint Conference on Computational Intelligence - ECTA, pages 197–204, 2024

    work page 2024

  3. [2]

    Alcaraz-Herrera, M.-A

    H. Alcaraz-Herrera, M.-A. Tsompanas, A. Adamatzky, and I. Balaz. Neuroevolution algorithms applied in the designing process of biohybrid actuators. https://arxiv.org/abs/2408.07671, 2024

    arXiv 2024

  4. [4]

    J. E. Auerbach and J. C. Bongard. Evolving complete robots with cppn-neat: The utility of recurrent connections. InProceedings of the 13th Annual Conference on Genetic and Evolutionary Computation. Association for Computing Machinery, 2011

    work page 2011

  5. [5]

    Blansché, P

    A. Blansché, P. Gançarski, and J. J. Korczak. Genetic algorithms for feature weighting: Evolution vs. coevolution and darwin vs. lamarck. InMICAI 2005: Advances in Artificial Intelligence, pages 682–691, 2005

    work page 2005

  6. [6]

    Hiller and H

    J. Hiller and H. Lipson. Dynamic simulation of soft multimaterial 3d-printed objects.Soft robotics, 1(1), 2014

    work page 2014

  7. [7]

    Kriegman

    S. Kriegman. Reconfigurable organisms. https://github.com/skriegman/reconfigurable_organisms. Accessed: 2025-01-10

    work page 2025

  8. [8]

    X. Ma, X. Li, Q. Zhang, K. Tang, Z. Liang, W. Xie, and Z. Zhu. A survey on cooperative co-evolutionary algorithms.IEEE Transactions on Evolutionary Computation, 23(3):421–441, 2019

    work page 2019

Show all 16 references
  1. [9]

    M. D. Schmidt and H. Lipson. Age-fitness pareto optimization. InProceedings of the 12th Annual Conference on Genetic and Evolutionary Computation, page 543–544, 2010

    2010

  2. [10]

    Schulz, C

    A. Schulz, C. Sung, A. Spielberg, W. Zhao, R. Cheng, E. Grinspun, D. Rus, and W. Matusik. Interactive robogami: An end-to-end system for design of robots with ground locomotion.The International Journal of Robotics Research, 36(10):1131–1147, 2016

    2016

  3. [11]

    K. O. Stanley. Compositional pattern producing networks: A novel abstraction of development.Genetic Programming and Evolvable Machines, 8(2):131–162, 2007

    2007

  4. [12]

    K. O. Stanley, D. B. D’Ambrosio, and J. Gauci. A Hypercube-Based Encoding for Evolving Large-Scale Neural Networks.Artificial Life, 15(2):185–212, 04 2009

    2009

  5. [13]

    K. O. Stanley and R. Miikkulainen. Evolving neural networks through augmenting topologies.Evolutionary Computation, 10(2):99–127, 2002

    2002

  6. [14]

    Tsompanas and I

    M.-A. Tsompanas and I. Balaz. Outline of an evolutionary morphology generator towards the modular design of a biohybrid catheter.Frontiers in Robotics and AI, 11:1337722, 2024

    2024

  7. [15]

    Wang and A

    J. Wang and A. Chortos. Control strategies for soft robot systems.Advanced Intelligent Systems, 4(5):88–101, 2022

    2022

  8. [16]

    R. Wen, Z. Guo, T. Zhao, X. Ma, Q. Wang, and Z. Wu. Neuroevolution of augmenting topologies based musculor-skeletal arm neurocontroller. In2017 IEEE International Instrumentation and Measurement Technology Conference (I2MTC), pages 1–6, 2017. 6

    2017

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.