pith. sign in

arxiv: 2606.11037 · v1 · pith:U4CYXQPMnew · submitted 2026-06-09 · 💻 cs.RO

Generation of Diverse and Functional Robot Designs using Superquadrics Parametrisation and Quality-Diversity

Leni Le Goff , Simon Smith , Emma Hart This is my paper

Pith reviewed 2026-06-27 13:08 UTC · model grok-4.3

classification 💻 cs.RO
keywords robot morphology generationsuperquadricsquality-diversity algorithmsMAP-Elitesevolutionary roboticsdesign space explorationmorphological diversity
0
0 comments X

The pith

Superquadrics parametrization combined with MAP-Elites produces the highest diversity and functionality scores for generated robot designs across two test environments.

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

The paper tests whether a compact geometric shape representation can help evolutionary methods avoid collapsing to similar robot bodies while still finding capable ones. It pits superquadrics, which describe 3D forms through a small set of tunable parameters, against compositional pattern producing networks as ways to create robot morphologies. These generators are paired either with ordinary evolutionary search or with the MAP-Elites quality-diversity algorithm that explicitly rewards both performance and behavioral difference. In the two environments examined, the superquadrics plus MAP-Elites pairing records the highest QD-score, meaning it returns both more varied and more effective designs than the other three combinations. The result matters because robot design spaces are enormous and most search methods quickly lose variety, leaving designers with few options to choose from.

Core claim

Superquadrics offer an interpretable and compact way to encode robot body geometry. When this representation is used to generate morphologies inside the MAP-Elites algorithm, the resulting collection of robots achieves the highest QD-score in both test environments, indicating greater coverage of distinct high-performing designs than is obtained with compositional pattern producing networks or with standard evolutionary algorithms.

What carries the argument

Superquadrics parametrization, a small set of mathematical parameters that define families of 3D shapes, used as the morphology generator inside the MAP-Elites quality-diversity algorithm.

If this is right

  • Designers obtain a larger set of distinct, functional robot bodies from a single run rather than repeated independent searches.
  • The compact parameter count of superquadrics reduces the dimensionality that the search algorithm must explore compared with network-based generators.
  • Explicit diversity maintenance becomes more effective when the underlying shape representation already constrains the space to geometrically meaningful variations.
  • Interpretable parameters make it easier for a human to understand or manually adjust the generated morphologies after the algorithm finishes.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same pairing could be tested on tasks that require multiple coordinated robots rather than single agents.
  • Because superquadrics are defined by continuous parameters, small mutations produce smooth shape changes that may preserve functionality better than discrete structural edits.
  • If fabrication constraints such as minimum feature size were added to the evaluation, the advantage of the compact representation might increase or decrease depending on how those constraints interact with the parameter bounds.

Load-bearing premise

The two chosen test environments and their fitness functions are representative enough of real robot design problems that performance differences observed there will hold outside the simulation.

What would settle it

Running the same four combinations on a third environment whose locomotion or manipulation demands differ markedly from the original two, and finding that superquadrics plus MAP-Elites no longer records the highest QD-score.

Figures

Figures reproduced from arXiv: 2606.11037 by Emma Hart, Leni Le Goff, Simon Smith.

Figure 1
Figure 1. Figure 1: ME2HKS : 1) Asynchronous Morpho-Evolution, EA to optimise robot designs; 2) a superquadrics-based encoding for the robot chassis (compared to CPPN); 3) Home￾okinetic controller, a self-organised behaviour generator; 4) MAP-Elites for diverse de￾signs. The output of ME2HKS is an archive of functional and diverse robot. components [5,27,4,16,20]. Due to the complexity and high-dimensionality of search spaces… view at source ↗
Figure 2
Figure 2. Figure 2: Five components for the robot’s design. Design Space The robot designs are based on the ARE robotic platform [3,1]. This platform utilises 5 hand-designed components (see [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: The two environments used in the experiments. From left to right image: flat and cracks. All the environments are 4 by 4 metres. A robot is included in the picture of the cracks environment, giving an idea of the relative size of the robots in respect to the environment. Compositional Producing Pattern Network (CPPN) The CPPN takes three in￾puts describing the spatial coordinates of the voxel and outputs o… view at source ↗
Figure 5
Figure 5. Figure 5: Pictures of robots picked from the top 10 robots in term of exploration score for ME2HKC and ME2HKS . See appendix E for more pictures. MAP-Elites, at the end of the evolutionary process the 10000 robots generated are assigned to the same 6-dimensional grid as used in MAP-Elites. The source code used to run the experiments and supplementary materials with the appendices are available on Zenodo. 4 Results O… view at source ↗
Figure 6
Figure 6. Figure 6: Distribution of the replicates over the max exploration score and archive cov￾erage [PITH_FULL_IMAGE:figures/full_fig_p011_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Distribution of the top 1% robots over joints and wheels. From top to bottom: MEHK - Dual CPPN, ME2HK - Dual CPPN, MEHK - SQ-CPPN, ME2HK - SQ-CPPN [PITH_FULL_IMAGE:figures/full_fig_p012_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Areas visited by the top 10 robots per replicates in term of exploration score using HKC. Each plot aggregates 200 robots trajectory. Finally, [PITH_FULL_IMAGE:figures/full_fig_p013_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Grid archive pair plot for each variant on the flat arena. The heat-maps show the projection of the archive on 2D and the histograms in the diagonal show the projection on one dimension, length of the bars is proportional to exploration score. Colours show exploration score: blue as 0%, red 25% and green 50% of the arena explored. in terms of quality and diversity, regardless of the algorithm used. This is… view at source ↗
read the original abstract

Generative design of robots requires navigating a vast search-space, encompassing physical configurations and behavioural parameters. Evolutionary Algorithms (EAs) have shown promising results, but often converge prematurely to a small set of sub-optimal designs. Most EAs fail to maintain sufficient diversity in the population that would allow the discovery of distinct functional robots. To counter premature convergence, we introduce a superquadrics-based representation (SQs) for robot bodies. SQs are interpretable, compact and computationally efficient mathematical representations of 3D geometrical shapes that can be tuned to specific design-spaces. To encourage morphological diversity, we combine this representation with a quality-diversity (QD) algorithm (MAP-Elites). We compare SQs and Compositional Pattern Producing Networks representations as generators of morphologies, combining them with standard EAs and MAP-Elites. In two test environments, we find that using SQs to generate morphology in conjunction with the MAP-Elites algorithm reaches the highest QD-score across both environments, maximising diversity of design and functionality of generated robots. The findings highlight the benefits of using a compact and interpretable geometric representation for exploring a complex design-space and suggest that combining SQs with an explicit diversity mechanism increases the quality and number of designs generated.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript proposes a superquadrics (SQs) parametrization for generating robot morphologies, combined with the MAP-Elites quality-diversity algorithm, to produce diverse and functional robot designs. It compares SQs against Compositional Pattern Producing Networks (CPPNs) as morphology generators, paired with both standard evolutionary algorithms and MAP-Elites, and reports that SQs + MAP-Elites achieves the highest QD-score across two test environments, thereby maximizing design diversity and functionality.

Significance. If the experimental comparisons hold under scrutiny, the work demonstrates that compact, interpretable geometric representations can improve exploration of robot design spaces when paired with explicit diversity mechanisms, offering a practical alternative to more complex generators like CPPNs for evolutionary robotics applications.

major comments (2)
  1. [Abstract and §4 (Experiments)] The headline result (highest QD-score for SQs + MAP-Elites) is load-bearing on the representativeness of the two test environments, their physics simulators, behavior descriptors, fitness functions, and MAP-Elites grid configuration; none of these are described in the abstract or methods overview, preventing assessment of whether the reported advantage is an artifact of the chosen testbed rather than a general property of the representation.
  2. [§3 (Methodology)] The comparison between SQs and CPPNs does not report the specific parameter counts, mutation operators, or initialization procedures used for each generator; without these details it is impossible to determine whether the performance difference arises from the representation itself or from unequal search-space dimensionality or operator bias.
minor comments (2)
  1. [Abstract] The abstract states that SQs are 'tuned to specific design-spaces' but provides no concrete bounds or constraints on the superquadric parameters; a short table or paragraph listing the ranges would improve reproducibility.
  2. [§5 (Results)] Figure captions and axis labels in the results section should explicitly state the number of independent runs and whether error bars represent standard deviation or standard error.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments. We address each major point below and will incorporate revisions to improve clarity and transparency.

read point-by-point responses

  1. Referee: [Abstract and §4 (Experiments)] The headline result (highest QD-score for SQs + MAP-Elites) is load-bearing on the representativeness of the two test environments, their physics simulators, behavior descriptors, fitness functions, and MAP-Elites grid configuration; none of these are described in the abstract or methods overview, preventing assessment of whether the reported advantage is an artifact of the chosen testbed rather than a general property of the representation.

    Authors: We agree that the abstract and methods overview should provide enough context on the experimental setup for readers to evaluate generalizability. While full details appear in §4, we will revise the abstract to briefly note the two environments and add a summary paragraph in the methods overview covering the simulators, behavior descriptors, fitness functions, and MAP-Elites grid. This will help address concerns about testbed specificity. revision: yes

  2. Referee: [§3 (Methodology)] The comparison between SQs and CPPNs does not report the specific parameter counts, mutation operators, or initialization procedures used for each generator; without these details it is impossible to determine whether the performance difference arises from the representation itself or from unequal search-space dimensionality or operator bias.

    Authors: We acknowledge that these implementation details are not explicitly stated in §3. To enable assessment of whether differences arise from the representations or from search-space or operator disparities, we will expand §3 with the parameter counts for SQs and CPPNs, the mutation operators and rates applied to each, and the initialization procedures. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical comparison of representations with no derivation or fitted-parameter reuse

full rationale

The paper reports an experimental comparison of superquadrics versus CPPN morphology generators, each paired with standard EAs and MAP-Elites, evaluated by QD-score in two test environments. No equations, parameter-fitting procedures, or predictive claims appear; the headline result is a direct empirical measurement of diversity and performance. No self-citation is invoked to justify a uniqueness theorem or ansatz, and no step reduces a reported outcome to its own inputs by construction. The work is therefore self-contained against external benchmarks and receives the default non-circularity finding.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract provides no explicit free parameters, axioms, or invented entities; superquadrics are treated as a standard mathematical tool.

pith-pipeline@v0.9.1-grok · 5753 in / 1004 out tokens · 17309 ms · 2026-06-27T13:08:18.013727+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

40 extracted references · 5 canonical work pages

  1. [1]

    Frontiers in Robotics and AI10(2023)

    Angus, M., Buchanan, E., Le Goff, L.K., Hart, E., Eiben, A.E., De Carlo, M., Winfield, A.F., Hale, M.F., Woolley, R., Timmis, J., et al.: Practical hardware for evolvable robots. Frontiers in Robotics and AI10(2023)

    2023

  2. [2]

    In: Proceedings of the 13th Annual Conference on Genetic and Evolutionary Computation

    Auerbach, J.E., Bongard, J.C.: Evolving complete robots with cppn-neat: the util- ity of recurrent connections. In: Proceedings of the 13th Annual Conference on Genetic and Evolutionary Computation. p. 1475–1482. GECCO ’11, Association for Computing Machinery, New York, NY, USA (2011).https://doi.org/10. 1145/2001576.2001775,https://doi.org/10.1145/200157...

    work page doi:10.1145/2001576.2001775 2011

  3. [3]

    Robotics9(4), 106 (2020)

    Buchanan, E., Le Goff, L.K., Li, W., Hart, E., Eiben, A.E., De Carlo, M., Winfield, A.F., Hale, M.F., Woolley, R., Angus, M., et al.: Bootstrapping artificial evolution to design robots for autonomous fabrication. Robotics9(4), 106 (2020)

    2020

  4. [4]

    Journal of The Royal Society In- terface15(143), 20170937 (2018)

    Cheney, N., Bongard, J., SunSpiral, V., Lipson, H.: Scalable co-optimization of morphology and control in embodied machines. Journal of The Royal Society In- terface15(143), 20170937 (2018)

    2018

  5. [5]

    ACM SIGEVOlution7(1), 11–23 (2014)

    Cheney, N., MacCurdy, R., Clune, J., Lipson, H.: Unshackling evolution: evolv- ing soft robots with multiple materials and a powerful generative encoding. ACM SIGEVOlution7(1), 11–23 (2014)

    2014

  6. [6]

    In: Proceedings of the 2020 Genetic and Evolutionary Computation Conference

    Colas, C., Madhavan, V., Huizinga, J., Clune, J.: Scaling map-elites to deep neu- roevolution. In: Proceedings of the 2020 Genetic and Evolutionary Computation Conference. pp. 67–75 (2020)

    2020

  7. [7]

    In: Proceedings of the Genetic and Evolution- ary Computation Conference

    Cully, A.: Multi-emitter map-elites: improving quality, diversity and data efficiency with heterogeneous sets of emitters. In: Proceedings of the Genetic and Evolution- ary Computation Conference. pp. 84–92 (2021)

    2021

  8. [8]

    VDI BERICHTE1679, 301–308 (2002)

    Der, R., Herrmann, J., Liebscher, R.: Homeokinetic approach to autonomous learn- ing in mobile robots. VDI BERICHTE1679, 301–308 (2002)

    2002

  9. [9]

    Der, R., Martius, G.: The playful machine: theoretical foundation and practical realization of self-organizing robots, vol. 15. Springer Science & Business Media (2012)

    2012

  10. [10]

    Dupac, M.: An object-oriented approach for mechanical components design and visualization. Eng. with Comput.28(2), 95–107 (Apr 2012)

    2012

  11. [11]

    Shirwatkar, N

    Goff, L.K.L., Smith, S.C.: Efficient and diverse generative robot designs using evolution and intrinsic motivation. In: 2025 IEEE International Conference on Robotics and Automation (ICRA). pp. 16598–16605 (2025).https://doi.org/ 10.1109/ICRA55743.2025.11127289

    work page doi:10.1109/icra55743.2025.11127289 2025

  12. [12]

    Nature communications12(1), 5721 (2021)

    Gupta, A., Savarese, S., Ganguli, S., Fei-Fei, L.: Embodied intelligence via learning and evolution. Nature communications12(1), 5721 (2021)

    2021

  13. [13]

    Journal of Intelligent & Robotic Systems109(2), 29 (2023)

    Han, X., Yang, L.: Sq-slam: Monocular semantic slam based on superquadric object representation. Journal of Intelligent & Robotic Systems109(2), 29 (2023)

    2023

  14. [14]

    Georg-August-Universitaet Goettingen (Germany) (2009)

    Hesse, F.: Self-Organizing Control for Autonomous Robots A Dynamical Systems Approach Based on the Principle of Homeokinesis. Georg-August-Universitaet Goettingen (Germany) (2009)

    2009

  15. [15]

    In: AISB workshop on evolutionary computing

    Husbands, P., Jermy, G., McIlhagga, M., Ives, R.: Two applications of genetic algorithms to component design. In: AISB workshop on evolutionary computing. pp. 50–61. Springer (1996)

    1996

  16. [16]

    Proceedings of the National Academy of Sciences 117(4), 1853–1859 (2020) 16 L

    Kriegman, S., Blackiston, D., Levin, M., Bongard, J.: A scalable pipeline for de- signing reconfigurable organisms. Proceedings of the National Academy of Sciences 117(4), 1853–1859 (2020) 16 L. K. Le Goff et al

    2020

  17. [17]

    In: Proceedings of the Genetic and Evolutionary Computation Conference Companion

    LeGoff,L.,Hart,E.:Improvingefficiencyofevolvingrobotdesignsviaself-adaptive learning cycles and an asynchronous architecture. In: Proceedings of the Genetic and Evolutionary Computation Conference Companion. pp. 1607–1615 (2024)

    2024

  18. [18]

    In: Artificial Life Conference Proceedings

    Le Goff, L.K., Hart, E., Coninx, A., Doncieux, S.: On pros and cons of evolving topologies with novelty search. In: Artificial Life Conference Proceedings. pp. 423–

  19. [19]

    MIT Press One Rogers Street, Cambridge, MA 02142-1209, USA (2020)

    2020

  20. [20]

    arXiv preprint arXiv:2411.18423 (2024)

    Le Goff, L.K., Smith, S.C.: Efficient and diverse generative robot designs using evolution and intrinsic motivation. arXiv preprint arXiv:2411.18423 (2024)

    arXiv 2024

  21. [21]

    IEEE Transactions on Evolutionary Computation28(6), 1561–1574 (2023)

    Li, W., Buchanan, E., Le Goff, L.K., Hart, E., Hale, M.F., Wei, B., De Carlo, M., Angus, M., Woolley, R., Gan, Z., et al.: Evaluation of frameworks that combine evolution and learning to design robots in complex morphological spaces. IEEE Transactions on Evolutionary Computation28(6), 1561–1574 (2023)

    2023

  22. [22]

    In: Linden- mayer systems: Impacts on theoretical computer science, computer graphics, and developmental biology, pp

    Lindenmayer, A., Jürgensen, H.: Grammars of development: discrete-state models for growth, differentiation, and gene expression in modular organisms. In: Linden- mayer systems: Impacts on theoretical computer science, computer graphics, and developmental biology, pp. 3–21. Springer (1992)

    1992

  23. [23]

    In: Proceedings of the Genetic and Evolutionary Computation Conference

    Medvet, E., Bartoli, A., Pigozzi, F., Rochelli, M.: Biodiversity in evolved voxel- based soft robots. In: Proceedings of the Genetic and Evolutionary Computation Conference. pp. 129–137 (2021)

    2021

  24. [24]

    In: European Conference on Genetic Programming (Part of EvoStar)

    Mertan, A., Cheney, N.: Investigating premature convergence in co-optimization of morphology and control in evolved virtual soft robots. In: European Conference on Genetic Programming (Part of EvoStar). pp. 38–55. Springer (2024)

    2024

  25. [25]

    In: Artificial Life Conference Proceedings 37

    Mertan, A., Cheney, N.: Evolutionary brain-body co-optimization consistently fails to select for morphological potential. In: Artificial Life Conference Proceedings 37. vol. 2025, p. 63. MIT Press One Rogers Street, Cambridge, MA 02142-1209, USA journals-info ... (2025)

    2025

  26. [26]

    Frontiers in Robotics and AI8, 672379 (2021)

    Miras, K.: Constrained by design: Influence of genetic encodings on evolved traits of robots. Frontiers in Robotics and AI8, 672379 (2021)

    2021

  27. [27]

    In: Artificial Life Conference Proceed- ings

    Miras, K., Haasdijk, E., Glette, K., Eiben, A.: Effects of selection preferences on evolved robot morphologies and behaviors. In: Artificial Life Conference Proceed- ings. pp. 224–231. MIT Press One Rogers Street, Cambridge, MA 02142-1209, USA (2018)

    2018

  28. [28]

    Mouret, J.B., Clune, J.: Illuminating search spaces by mapping elites (2015), https://arxiv.org/abs/1504.04909

    Pith/arXiv arXiv 2015

  29. [29]

    In: Genetic Programming: 28th European Conference, EuroGP 2025, Held as Part of EvoStar 2025, Tri- este, Italy, April 23–25, 2025, Proceedings

    Nadizar, G., Rusin, F., Medvet, E., Ochoa, G.: The role of stepping stones in map-elites: Insights from search trajectory networks. In: Genetic Programming: 28th European Conference, EuroGP 2025, Held as Part of EvoStar 2025, Tri- este, Italy, April 23–25, 2025, Proceedings. p. 224–239. Springer-Verlag, Berlin, Heidelberg (2025).https://doi.org/10.1007/97...

    work page doi:10.1007/978-3-031-89991-1_14 2025

  30. [30]

    Frontiers in Robotics and AI 8, 639173 (2021)

    Nordmoen, J., Veenstra, F., Ellefsen, K.O., Glette, K.: Map-elites enables powerful stepping stones and diversity for modular robotics. Frontiers in Robotics and AI 8, 639173 (2021)

    2021

  31. [31]

    In: Proceedings of the Genetic and Evolutionary Compu- tation Conference

    Scarton, L., Hagg, A.: On the suitability of representations for quality diversity optimization of shapes. In: Proceedings of the Genetic and Evolutionary Compu- tation Conference. p. 963–971. GECCO ’23, Association for Computing Machin- ery, New York, NY, USA (2023).https://doi.org/10.1145/3583131.3590381, https://doi.org/10.1145/3583131.3590381

    work page doi:10.1145/3583131.3590381 2023

  32. [32]

    In: IAPR In- ternational Workshop on Partially Supervised Learning

    Smith, S.C., Herrmann, J.M.: Homeokinetic reinforcement learning. In: IAPR In- ternational Workshop on Partially Supervised Learning. pp. 82–91. Springer (2011) Generation of Diverse and Func. Robot using SQ Parametrisation and QD 17

    2011

  33. [33]

    Science201(4356), 579–582 (1978)

    Soodak, H., Iberall, A.: Homeokinetics: A physical science for complex systems. Science201(4356), 579–582 (1978)

    1978

  34. [34]

    Genetic programming and evolvable machines8, 131–162 (2007)

    Stanley, K.O.: Compositional pattern producing networks: A novel abstraction of development. Genetic programming and evolvable machines8, 131–162 (2007)

    2007

  35. [35]

    In: Artificial Life Conference Proceedings 37

    Thomson, S.L., Le Goff, L.K., Hart, E., Eiben, A.E., Luck, K.S.: Understanding the navigation of robot morphology spaces with local optima network analysis. In: Artificial Life Conference Proceedings 37. vol. 2025, p. 42. MIT Press One Rogers Street, Cambridge, MA 02142-1209, USA journals-info ... (2025)

    2025

  36. [36]

    In: Proceedings of the Genetic and Evolutionary Computation Conference Companion

    Veenstra, F., Olsen, M.H., Glette, K.: Effects of encodings and quality-diversity on evolving 2d virtual creatures. In: Proceedings of the Genetic and Evolutionary Computation Conference Companion. p. 164–167. GECCO ’22, Association for Computing Machinery, New York, NY, USA (2022).https://doi.org/10.1145/ 3520304.3529053,https://doi.org/10.1145/3520304.3529053

    work page doi:10.1145/3520304.3529053 2022

  37. [37]

    In: 2017 IEEE International Conference on Robotics and Automation (ICRA)

    Vezzani, G., Pattacini, U., Natale, L.: A grasping approach based on superquadric models. In: 2017 IEEE International Conference on Robotics and Automation (ICRA). pp. 1579–1586. IEEE (2017)

    2017

  38. [38]

    Autonomous Robots49(3), 1–16 (2025)

    Wu, Y., Li, W., Liu, Z., Liu, W., Chirikjian, G.S.: Autonomous learning-free grasp- ing and robot-to-robot handover of unknown objects. Autonomous Robots49(3), 1–16 (2025)

    2025

  39. [39]

    arXiv preprint arXiv:2503.16127 (2025)

    Xie, Y., Chu, K.f., Wang, X., Iida, F.: The morphology-control trade-off: Insights into soft robotic efficiency. arXiv preprint arXiv:2503.16127 (2025)

    arXiv 2025

  40. [40]

    In: 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)

    Xie, Y., Pinskier, J., Liow, L., Howard, D., Iida, F.: A’map’to find high-performing soft robot designs: Traversing complex design spaces using map-elites and topology optimization. In: 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 11408–11415. IEEE (2024)

    2024