Pith. sign in

REVIEW 4 major objections 5 minor 37 references

NEAT and HyperNEAT based Design for Soft Actuator Controllers

T0 review · 4 major / 5 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 NEAT and HyperNEAT outperform a standard genetic algorithm at designing phase-offset controllers for simulated soft actuator morphologies, with NEAT the best of the three.

desk verdict Incremental but statistically careful empirical study; the NEAT-over-HyperNEAT claim is undermined by a generic 2D substrate that the authors themselves concede disables HyperNEAT's geometry mechanism. read the letter →

arxiv 2506.04698 v1 pith:Y7B7ZQQH submitted 2025-06-05 cs.NE

Hugo Alcaraz-Herrera , Michail-Antisthenis Tsompanas , Igor Balaz , Andrew Adamatzky This is my paper
classification cs.NE
keywords NEATHyperneuroevolutionsoftroboticsCPPNactuatorcontrollerVoxelyzephaseoffset
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 asks whether neuroevolution that grows network topology can design better controllers for soft actuators than a fixed-parameter genetic search. It compares NEAT and HyperNEAT against a standard genetic algorithm on the task of finding phase offsets for the active voxels of simulated catheter-like actuators so that the free end bends upward as far as possible. The central claim is that both neuroevolution methods learn faster and reach higher bending displacement, and that NEAT is the best: across all tested morphologies and activation-function dictionaries, the ranking is NEAT > HyperNEAT > SGA with statistically significant differences. If true, this indicates that evolving both the controller topology and its weights, rather than optimizing a fixed array of phase values, is a more practical route to controlling compliant actuators.

What carries the argument

The load-bearing object is the Compositional Pattern-Producing Network (CPPN), used as a controller: it takes a voxel's coordinates $(x_i, y_i, z_i)$ and material type $m_i$ as input and outputs that voxel's phase offset $pho_i$, clamped to $[-2\pi, 2\pi]$. Under NEAT the CPPN's own topology and weights are evolved from minimal networks by adding nodes, connections, and activation functions; under HyperNEAT the CPPN instead generates the weights and biases of a separately fixed substrate network. Fitness comes from a voxel-based physics engine (Voxelyze) that measures upward bending displacement in the $yz$ plane, with material parameters taken from prior in vitro work. This lets evolution search directly over control patterns that are tied to morphology and material placement.

What would settle it

Build or reproduce a catheter-like soft actuator from the simulated morphologies, run the NEAT-, HyperNEAT-, and SGA-designed phase-offset controllers on it, and measure upward bending: if NEAT does not exceed SGA on the physical device, or if the controllers do not transfer at all, the claimed ranking fails.

Watch

Extended reading notes

Core claim

The paper's central discovery is that neuroevolution of augmenting topologies (NEAT) is the most effective of the three tested methods for designing phase-offset controllers for soft actuator morphologies (SAMs). NEAT and HyperNEAT both outperform a standard genetic algorithm, and NEAT in turn outperforms HyperNEAT, a result that holds for the nine fittest and nine worst morphologies and for both a full and a reduced set of activation functions. The differences are statistically significant, with Kruskal-Wallis and Dunn's tests giving $p<0.01$. NEAT also yields markedly simpler controllers, with a mean of 2-3 hidden nodes and 2-3 connections, whereas HyperNEAT's fixed substrate produces 33-43 connections and 13 hidden nodes. The paper attributes NEAT's advantage to its unrestricted search over topology, weights, and activation functions, and attributes HyperNEAT's weaker showing to a fixed two-dimensional substrate that cannot embody the geometry of the three-dimensional soft actuators.

Load-bearing premise

The load-bearing premise is that the voxel-based physics simulation, with material parameters taken from prior in vitro work, faithfully represents the mechanical behavior of real soft actuators; the paper does not validate against a physical actuator.

Editorial extensions

If this is right

  • Controllers evolved by NEAT reached near-optimal bending within the first generations, so neuroevolution can reduce the number of simulation evaluations needed compared with SGA.
  • The ranking NEAT > HyperNEAT > SGA holds on both easy and hard morphologies, suggesting the advantage is not specific to well-formed actuators.
  • NEAT's resulting networks are small enough (2-3 connections) to be plausibly implemented on real embedded hardware.
  • HyperNEAT's fixed substrate limits it; the paper implies that a substrate able to reflect 3D morphology might close the gap.
  • The full activation-function dictionary gives statistically significant but practically small improvements, so reduced dictionaries may be preferable in practice.

Reading between the lines

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

  • If the observed advantage transfers to hardware, NEAT could serve as the controller-design step in an automated pipeline that co-optimizes morphology and control for catheters or other compliant medical devices.
  • Because the gap between full and reduced activation dictionaries is small, a production implementation could use the reduced dictionary to shorten search time with little loss.
  • A direct testable extension is to check whether HyperNEAT improves once the substrate is replaced by a three-dimensional layout matched to the actuator; the paper itself identifies the fixed 2D substrate as its likely weakness.
  • The paper's reliance on simulation means the ranking is about simulated actuators; establishing the same ordering on physical soft actuators is the next step.
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

4 major / 5 minor

Summary. The paper proposes using NEAT and HyperNEAT to evolve phase-offset controllers for simulated soft actuator morphologies (SAMs) in the Voxelyze physics engine, comparing them against a standard genetic algorithm (SGA). Experiments use nine fittest and nine worst SAMs from the authors' prior work, two activation-function dictionaries (full and reduced), and 20 evolutionary runs per configuration. Four metrics are reported: general performance (upward bending displacement), robustness across SAMs, controller complexity, and activation-function usage. The central claims are that NEAT and HyperNEAT outperform SGA in all tested scenarios, and that NEAT performs best overall (NEAT > HyperNEAT > SGA, Dunn's test p < 0.01), while producing simpler networks.

Significance. If the results are correct, the paper provides evidence that topology-growing neuroevolution, especially NEAT, can design effective and simple phase-offset controllers for simulated soft actuators, with potential relevance to catheter-like medical devices. The study's strengths include the use of multiple statistical tests (Shapiro-Wilk, Kruskal-Wallis, Dunn, Wilcoxon), 95% confidence intervals, 20 repeated runs, and a comparison across two morphology groups and two activation-function dictionaries. The main significance is contingent on the simulation faithfully representing real soft-actuator behavior and on the HyperNEAT comparison being a fair algorithmic test, both of which are questionable as discussed below.

major comments (4)
  1. [§4.2.2, Fig. 3, §6] The NEAT > HyperNEAT ranking is load-bearing for the paper's central claim, but the HyperNEAT substrate is a generic two-dimensional grid with a fixed 13-node hidden layer, and the authors explicitly concede in Section 6 that 'no geometrical aspects of the domain problem are available that can be embodied in the design of the substrate' and that the substrate was designed 'only considering the allocation of the neurons in a two-dimension space.' Since HyperNEAT's principal advantage over plain NEAT is its ability to exploit domain geometry through CPPN-generated connectivity, this configuration choice handicaps HyperNEAT in a way that could reverse the ranking under a more appropriate 3D substrate or ES-HyperNEAT. The abstract's unconditional statement that 'NEAT performed better for all different scenarios tested' is therefore not supported as a general algorithmic comparison; it is at most a statement about this particular generic-substrate implementation. The authors should either re-run HyperNEAT with a domain-informed substrate or substantially temper the claim.
  2. [§5.1, Figs. 5–8] No exact numerical results are reported anywhere in the paper. All performance claims rest on figures with shaded confidence intervals and on p-values reported only as 'p < 0.01', with no test statistics, effect sizes, medians, interquartile ranges, or sample sizes beyond '20 duplicate runs'. This makes it impossible for a reader to verify the magnitude of the reported differences, which the text itself describes as 'minimal' and 'quite small'. The paper should report numerical summaries (e.g., final-generation median displacement with IQR for each algorithm/dictionary/SAM group) and effect sizes for the Dunn and Wilcoxon tests, or make the raw data available.
  3. [§4.1] The entire evaluation rests on the Voxelyze simulation with material parameters taken from prior in vitro work (Poisson's ratio 0.35, Young's modulus 5e6 Pa, ±50% volumetric actuation at 4 Hz), but the paper provides no validation that this simulation reproduces the bending behavior of real soft actuators in the relevant regime. The conclusion that NEAT is 'more appropriate' and 'easier to implement in real life applications' is therefore only as strong as the simulation's fidelity. A sensitivity analysis over the material parameters, or at least a discussion of how parameter uncertainty could affect the algorithm ranking, would be needed to support the practical claims.
  4. [§5.2, Eq. (8)] The robustness metric is not fully specified. Equation 8 averages displacement over nine SAMs for a 'controller h', but the text does not clearly explain whether the fittest controller for each SAM was evolved separately and then tested on all nine, or whether one controller was evolved against all nine SAMs simultaneously. The description 'utilising NF across 20 evolutionary trials' is ambiguous, and the violin plots in Figs. 9 and 10 do not clarify the unit of analysis. This ambiguity affects the interpretation of the robustness ranking and should be resolved with an explicit description of the experimental protocol.
minor comments (5)
  1. [§5.4, paragraph after Fig. 11] In the description of NEAT with the reduced dictionary on the nine fittest SAMs, the text says 'in a similar fashion to the scenario when NW is used', which appears to be a typo for 'when NF is used'; the surrounding text is about NF results.
  2. [Throughout] There are several typographical and formatting issues, including 'V oxelyze' (Section 4.1), 'Shapiro-Will' (Section 5.2), and inconsistent spacing around references. A careful proofreading pass is recommended.
  3. [Table 2] The mean number of connections and hidden nodes is reported without standard deviations or ranges, even though these are averages over 20 runs; reporting variability would strengthen the complexity comparison.
  4. [References] Reference [10] lists the authors as 'Haasdijk, Evertand Rusu' and 'A. E. Eiben'; the author names appear corrupted and should be corrected to 'E. Haasdijk, A. Rusu, and A. E. Eiben'.
  5. [§5.1] The statement 'the difference between NEAT and HyperNEAT performances is not visible within the y-axis' (Fig. 5) suggests the figure may not clearly show the claimed significant difference; a zoomed panel or numerical annotation would help the reader see the effect.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the NEAT-vs-HyperNEAT-vs-SGA ranking is produced by new evolutionary runs in simulation, not by a fitted quantity or self-cited theorem.

full rationale

The paper's central claim—NEAT outperforms HyperNEAT and SGA across four scenarios—is an empirical outcome of newly executed evolution runs evaluated by Voxelyze displacement, not a quantity fitted from target data. No equation in the paper defines the controller output (phase offsets, Eqs. 6 and 7) in terms of the fitness metric (yz-plane displacement); fitness is maximized during evolution as described in Section 5.1. The SGA baseline is independently implemented with a phase-offset matrix, so the comparison is a fresh benchmark. The reuse of self-cited material ([2], [3], [4]) supplies testbed morphologies, hyperparameters (Table 1), and the HyperNEAT substrate (Fig. 3), but these are inputs to the experiment, not conclusions derived from them. The paper's admission in Section 6 that the substrate lacks domain geometry is a limitation and a potential fairness concern for the HyperNEAT comparison, but it is an implementation choice, not a circular reduction: the NEAT-vs-HyperNEAT result is still measured, not assumed. There is no 'prediction' that is statistically forced, no uniqueness theorem imported from the authors, and no ansatz presented as derived from prior work. The absence of physical validation of Voxelyze is a correctness/validity risk, not circularity. Accordingly, no load-bearing step reduces to its own inputs.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

This is an empirical comparison paper, so the free parameters are evolution hyperparameters and hand-set design choices rather than physically fitted constants. The central claim rests on the simulator's fidelity and on the fairness of the chosen substrate design, which are listed as domain assumptions.

free parameters (6)
  • Population size and number of generations = 50 individuals, 200 generations
    Set in Section 4.3 following previous work [3]; not tuned in this paper.
  • NEAT/HyperNEAT evolution hyperparameters = compatibility threshold 3; disjoint coefficient 1.0; weight coefficient 0.5; max stagnation 25; survival threshold…
    Taken from Table 1, which is copied from [3]; no sensitivity analysis reported.
  • HyperNEAT connection presence threshold and weight range = threshold 0.2; normalization to [-3.0, 3.0]
    Set in Section 4.2.2 by hand; no sensitivity analysis.
  • Substrate neuron allocation = two-dimensional substrate; number of layers and neurons per layer chosen in the [3,10] range by 'a series of…
    Section 4.2.2 states the choice was experimentally determined but does not describe the experiments or the optimization criterion.
  • Activation function dictionaries = FD: 23 functions including Sine, Gaussian, Inverse, etc.; RD: 9 functions (Sine, Negative sine, Squared, Negative…
    Manually curated in Section 4.3 to test whether a larger function set changes controller quality; the dictionaries are a controlled experimental variable, not fitted.
  • SGA crossover and mutation probabilities = crossover probability 0.9; mutation probability 0.1; real-valued matrix encoding in [-2π, 2π]
    Set in Section 5 and not tuned.
assumptions (5)
  • domain assumption Voxelyze simulator accurately models the mechanical response of soft actuators, including volumetric actuation of active voxels (±50% at 4 Hz).
    All fitness evaluations are based on this simulation (Section 4.1); no physical validation is performed in this paper.
  • domain assumption Material parameters (Poisson's ratio 0.35, Young's modulus 5e6 Pa, friction coefficients 1.0 and 0.5) from prior in vitro work apply to this design.
    Section 4.1 states these are based on previous research [15]; they are inputs to the physics engine.
  • domain assumption Encoding a controller as a CPPN queried by voxel coordinates and material type, with a phase-offset output, is an adequate representation space for SAM control.
    Sections 4.2.1 and 4.2.2 define the input/output encoding; no comparison to other controller representations is provided.
  • domain assumption Upward bending displacement in the yz plane is an appropriate fitness metric for a drug-delivering catheter.
    Section 4.1 selects this objective based on the application; no clinical validation is given.
  • ad hoc to paper A standard genetic algorithm is a representative baseline for 'traditional approaches'.
    SGA is introduced in Section 5 as the only baseline; the conclusions generalize to 'traditional approaches' from this single baseline.

how reviews work

0 comments
Cite this review

Pith. "Pith review of NEAT and HyperNEAT based Design for Soft Actuator Controllers." pith.science (2026). https://pith.science/paper/Y7B7ZQQH

@misc{pith2026250604698,
  author       = {Pith},
  title        = {Pith review of: NEAT and HyperNEAT based Design for Soft Actuator Controllers},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Y7B7ZQQH}},
  note         = {Machine review of arXiv:2506.04698}
}
read the original abstract

Since soft robotics are composed of compliant materials, they perform better than conventional rigid robotics in specific fields, such as medical applications. However, the field of soft robotics is fairly new, and the design process of their morphology and their controller strategies has not yet been thoroughly studied. Consequently, here, an automated design method for the controller of soft actuators based on Neuroevolution is proposed. Specifically, the suggested techniques employ Neuroevolution of Augmenting Topologies (NEAT) and Hypercube-based NEAT (HyperNEAT) to generate the synchronization profile of the components of a simulated soft actuator by employing Compositional Pattern Producing Networks (CPPNs). As a baseline methodology, a Standard Genetic Algorithm (SGA) was used. Moreover, to test the robustness of the proposed methodologies, both high- and low-performing morphologies of soft actuators were utilized as testbeds. Moreover, the use of an affluent and a more limited set of activation functions for the Neuroevolution targets was tested throughout the experiments. The results support the hypothesis that Neuroevolution based methodologies are more appropriate for designing controllers that align with both types of morphologies. In specific, NEAT performed better for all different scenarios tested and produced more simplistic networks that are easier to implement in real life applications.

Figures

Figures reproduced from arXiv: 2506.04698 by the authors.

Figure 1
Figure 1. Example of a genotype and its phenotype under NEAT. Figure adapted from [27]. [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. An example of a two-dimensional substrate configuration known as [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Substrate utilised under HyperNEAT to design SAM controllers. Image taken from [3]. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: Examples of SAMs found in [4]: (a) fittest and (b) worst. [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: Nine fittest SAMs: Mean general performance with [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]
Figure 6
Figure 6. Figure 6: Nine fittest SAMs: Left - Mean general performance with [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]
Figure 7
Figure 7. Figure 7: Nine worst SAMs: Mean general performance with [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]
Figure 8
Figure 8. Figure 8: Nine worst SAMs: Left - Mean general performance with [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]
Figure 9
Figure 9. Figure 9: Nine fittest SAMs: Left - Displacement observed in the [PITH_FULL_IMAGE:figures/full_fig_p012_9.png]
Figure 10
Figure 10. Figure 10: Nine worst SAMs: Left - Displacement observed in the [PITH_FULL_IMAGE:figures/full_fig_p013_10.png]
Figure 11
Figure 11. Figure 11: Nine fittest SAMs: Activation functions employed to compose the fittest controllers found by NEAT using: [PITH_FULL_IMAGE:figures/full_fig_p014_11.png]
Figure 12
Figure 12. Figure 12: Nine worst SAMs: Activation functions employed to compose the fittest controllers found by NEAT using: [PITH_FULL_IMAGE:figures/full_fig_p015_12.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

37 extracted references · 37 canonical work pages

  1. [1]

    Alcaraz-Herrera and J

    H. Alcaraz-Herrera and J. Cartlidge. Exploration of ontological representations for evolutionary computation. In 2022 IEEE Congress on Evolutionary Computation (CEC), pages 1–8, 2022

    work page 2022

  2. [2]

    NeuroEvolution algorithms applied in the designing process of biohybrid actuators

    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

    work page Pith review arXiv 2024

  3. [3]

    Alcaraz-Herrera, M.-A

    H. Alcaraz-Herrera, M.-A. Tsompanas, I. Balaz, and A. Adamatzky. Control of biohybrid actuators using neuroevolution. In Proceedings of the 16th International Joint Conference on Computational Intelligence, page 197–204, 2024

    work page 2024

  4. [4]

    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

  5. [5]

    Caceres, J

    C. Caceres, J. M. Rosario, and D. Amaya. Approach of kinematic control for a nonholonomic wheeled robot using artificial neural networks and genetic algorithms. In 2017 International Conference and Workshop on Bioinspired Intelligence (IWOBI), pages 1–6, 2017

    work page 2017

  6. [6]

    Drchal, J

    J. Drchal, J. Koutnik, and M. Snorek. Hyperneat controlled robots learn how to drive on roads in simulated environment. In 2009 IEEE Congress on Evolutionary Computation, pages 1087–1092, 2009

    work page 2009

  7. [7]

    Glorot, A

    X. Glorot, A. Bordes, and Y . Bengio. Deep sparse rectifier neural networks. In Proceedings of the Fourteenth International Conference on Artificial Intelligence and Statistics, Proceedings of Machine Learning Research, pages 315–323, 2011

    work page 2011

  8. [8]

    D. E. Goldberg and J. Richardson. Genetic algorithms with sharing for multimodal function optimization. In Proceedings of the Second International Conference on Genetic Algorithms on Genetic Algorithms and Their Application, page 41–49, 1987

    work page 1987

Show all 37 references
  1. [9]

    Gupta, S

    A. Gupta, S. Savarese, S. Ganguli, and L. Fei-Fei. Embodied intelligence via learning and evolution. Nature communications, 12(1):5721, 2021

    2021

  2. [10]

    A. A. Haasdijk, Evertand Rusu and A. E. Eiben. Hyperneat for locomotion control in modular robots. InEvolvable Systems: From Biology to Hardware, pages 169–180, 2010

    2010

  3. [11]

    Hiller and H

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

    2014

  4. [12]

    Hsiao, J.-Y

    J.-H. Hsiao, J.-Y . Chang, and C.-M. Cheng. Soft medical robotics: clinical and biomedical applications, challenges, and future directions. Advanced Robotics, 33(21):1099–1111, 2019. 16 A PREPRINT - AUGUST 16, 2025

    2019

  5. [13]

    Huang, J

    P.-C. Huang, J. Lehman, A. K. Mok, R. Miikkulainen, and L. Sentis. Grasping novel objects with a dexterous robotic hand through neuroevolution. In 2014 IEEE Symposium on Computational Intelligence in Control and Automation (CICA), pages 1–8, 2014

    2014

  6. [14]

    Kriegman

    S. Kriegman. Reconfigurable organisms. https://github.com/skriegman/reconfigurable_organisms. Accessed: 2025-03-01

    2025

  7. [15]

    Kriegman, D

    S. Kriegman, D. Blackiston, M. Levin, and J. Bongard. A scalable pipeline for designing reconfigurable organisms. Proceedings of the National Academy of Sciences, 117(4):1853–1859, 2020

    2020

  8. [16]

    E. Milana. Soft robotics for infrastructure protection. Frontiers in Robotics and AI, 9, 2022

    2022

  9. [17]

    Ororbia, A

    A. Ororbia, A. ElSaid, and T. Desell. Investigating recurrent neural network memory structures using neuro- evolution. In Proceedings of the Genetic and Evolutionary Computation Conference, GECCO ’19, page 446–455, 2019

    2019

  10. [18]

    S.-J. Park, M. Gazzola, K. S. Park, S. Park, V . Di Santo, E. L. Blevins, J. U. Lind, P. H. Campbell, S. Dauth, A. K. Capulli, et al. Phototactic guidance of a tissue-engineered soft-robotic ray. Science, 353(6295):158–162, 2016

    2016

  11. [19]

    Risi and K

    S. Risi and K. O. Stanley. An Enhanced Hypercube-Based Encoding for Evolving the Placement, Density, and Connectivity of Neurons. Artificial Life, 18(4):331–363, 10 2012

    2012

  12. [20]

    Risi and K

    S. Risi and K. O. Stanley. Confronting the challenge of learning a flexible neural controller for a diversity of morphologies. In Proceedings of the 15th Annual Conference on Genetic and Evolutionary Computation, GECCO ’13, page 255–262, 2013

    2013

  13. [21]

    G. D. Ruebsamen. Evolving intelligent embodied agents within a physically accurate environment. California State University, Long Beach, 2002

    2002

  14. [22]

    Rus and M

    D. Rus and M. T. Tolley. Design, fabrication and control of soft robots. Nature, 521(7553):467–475, 2015

    2015

  15. [23]

    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

  16. [24]

    Seriani, L

    S. Seriani, L. Marcini, M. Caruso, P. Gallina, and E. Medvet. Crowded environment navigation with neat: Impact of perception resolution on controller optimization. Journal of Intelligent & Robotic Systems, 101(2), 2021

    2021

  17. [25]

    Sokolov, I

    M. Sokolov, I. Afanasyev, A. Klimchik, and N. Mavridis. Hyperneat-based flipper control for a crawler robot motion in 3d simulation environment. In 2017 IEEE International Conference on Robotics and Biomimetics (ROBIO), pages 2652–2656, 2017

    2017

  18. [26]

    Song, W.-Y

    Y .-L. Song, W.-Y . Chiu, and S. H. Manoharan. Mixed compositional pattern-producing network-neuroevolution of augmenting topologies method for the locomotion control of a snake-like modular robot. In 2023 SICE International Symposium on Control Systems (SICE ISCS), pages 93–100, 2023

    2023

  19. [27]

    Stanley and R

    K. Stanley and R. Miikkulainen. Efficient evolution of neural network topologies. In Proceedings of the 2002 Congress on Evolutionary Computation. CEC’02 (Cat. No.02TH8600), volume 2, pages 1757–1762 vol.2, 2002

    2002

  20. [28]

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

    2007

  21. [29]

    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

  22. [30]

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

    2002

  23. [31]

    Tanaka and C

    F. Tanaka and C. Aranha. Co-evolving morphology and control of soft robots using a single genome. In 2022 IEEE Symposium Series on Computational Intelligence (SSCI), pages 1235–1242, 2022

    2022

  24. [32]

    A. S. Thakur, A. B. Awari, Z. Lyu, and T. Desell. Efficient neuroevolution using island repopulation and simplex hyperparameter optimization. In 2023 IEEE Symposium Series on Computational Intelligence (SSCI) , pages 1837–1842, 2023

    2023

  25. [33]

    Tibermacine and N

    A. Tibermacine and N. Djedi. Neat neural networks to control and simulate virtual creature’s locomotion. In 2014 International Conference on Multimedia Computing and Systems (ICMCS), pages 9–14, 2014

    2014

  26. [34]

    Tsompanas

    M.-A. Tsompanas. Biomeld reconfigurable organisms. https://github.com/Antisthenis/ reconfigurable_organisms/tree/biomeld_dev2. Accessed: January 10, 2025

    2025

  27. [35]

    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, 2024. 17 A PREPRINT - AUGUST 16, 2025

    2024

  28. [36]

    E. C. R. G. (UCF). The hybercube-based neuroevolution of augmenting topologies (hyperneat) users page. http://eplex.cs.ucf.edu/hyperNEATpage/, 2016. Accessed: January 10, 2025

    2016

  29. [37]

    Wang and A

    J. Wang and A. Chortos. Control strategies for soft robot systems. Advanced Intelligent Systems, 4(5):88–101, 2022. 18

    2022

Pith tools

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