pith. sign in

arxiv: 2605.15769 · v1 · pith:ENWIUDW2new · submitted 2026-05-15 · 💻 cs.RO · cs.AI

Lamarckian Inheritance in Dynamic Environments: How Key Variables Affect Evolutionary Dynamics

K. Ege de Bruin , Kyrre Glette , Kai Olav Ellefsen This is my paper

Pith reviewed 2026-05-20 18:47 UTC · model grok-4.3

classification 💻 cs.RO cs.AI
keywords Lamarckian inheritanceevolutionary roboticsdynamic environmentsmorphology optimizationlifetime learningBayesian optimizationreinforcement learning
0
0 comments X

The pith

Lamarckian inheritance in robot evolution only underperforms Darwinian when environmental changes are both conflicting and unpredictable.

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

The paper investigates why Lamarckian inheritance sometimes helps and sometimes hurts in evolving robots in changing environments. It establishes that the performance advantage of passing learned controller parameters to offspring depends on whether the environment changes require completely different control strategies and whether those changes can be anticipated. Experiments with soft robots using two learning methods show that Lamarckian inheritance is beneficial unless changes are conflicting and unpredictable at the same time. When a sensor is added to detect the change, the benefit returns because the robot can switch behaviors accordingly.

Core claim

In dynamic environments for co-optimizing robot morphology and control, Lamarckian inheritance of learned controllers only underperforms standard Darwinian evolution when the environmental changes are both conflicting—meaning they demand incompatible control strategies—and unpredictable for the agent. This is demonstrated across Bayesian optimization and reinforcement learning controllers in virtual soft robots, and the addition of an environmental change sensor restores Lamarckian advantages in conflicting cases by enabling prediction of the required behavioral shift.

What carries the argument

The dependence of Lamarckian inheritance benefits on the conflict level and predictability of environmental changes, where conflict means changes that invalidate previous control and predictability allows anticipation of the change.

If this is right

  • Lamarckian inheritance improves performance in dynamic environments that are either non-conflicting or predictable.
  • Adding a sensor for detecting changes allows Lamarckian inheritance to succeed even in conflicting dynamic environments.
  • The choice of learning method does not alter the dependence on conflict and predictability.
  • The results explain conflicting findings in prior evolutionary robotics literature on Lamarckian inheritance.

Where Pith is reading between the lines

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

  • Robotic systems in real-world varying conditions could incorporate change-detection sensors to leverage lifetime learning inheritance.
  • Testing the same variables in physical robot experiments would strengthen generalization beyond simulation.
  • These findings may extend to other domains where agents adapt controllers in changing tasks.

Load-bearing premise

The specific simulated dynamic environments and the two learning methods used are representative of broader settings in evolutionary robotics.

What would settle it

A demonstration that Lamarckian inheritance underperforms even in non-conflicting or predictable environments, or fails to benefit from a change-detection sensor in conflicting cases, would falsify the central claim.

Figures

Figures reproduced from arXiv: 2605.15769 by Kai Olav Ellefsen, K. Ege de Bruin, Kyrre Glette.

Figure 1
Figure 1. Figure 1: We classify dynamic environments along two dimensions: environmental conflict (whether optimal control is con [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the robot brain and its role in re [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Example of a rugged environment. Percentage [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Mean performance of the population over gener [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: The bidirectional environment. It is a flat environ [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Mean performance of the population over gener [PITH_FULL_IMAGE:figures/full_fig_p006_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Mean performance of the inherited brain, i.e. be￾fore learning, of the population over generations in a bidirec￾tional environment, both with and without direction sensor. Shaded areas indicate interquartile range (25th to 75th per￾centile). The lines are smoothened with a window size of 2 generations. Conflicting environment change Now that we have shown a case where the environment in not conflicting, an… view at source ↗
read the original abstract

The co-optimization of a robot's body and brain presents a coupled challenge: the morphology constrains which control strategies are effective, while the control determines how well the morphology performs. To address this, we combine morphology optimization as evolution with controller optimization as lifetime learning, utilizing Lamarckian inheritance to transfer learned controller parameters from parent to offspring. In dynamic environments, existing literature presents conflicting evidence: while traditional evolutionary theory often suggests Lamarckian inheritance lacks benefit, recent studies in evolutionary robotics indicate it can improve performance. We hypothesize that this is because previous works have not included all relevant variables with dynamic environments. In this work, we show that the benefit of Lamarckian inheritance depends on two variables: how conflicting the environmental changes are to robot control, and the predictability of those changes for the robotic agent. Using virtual soft robots and two different learning approaches, Bayesian optimization and reinforcement learning, we show that Lamarckian inheritance only underperforms Darwinian inheritance when the changes are both conflicting and unpredictable. We find that adding a sensor to detect environmental changes restores the benefits for Lamarckian inheritance in conflicting environments, by allowing robotic agents to predict the need for a different behavior, thereby generalizing their control.

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 investigates Lamarckian versus Darwinian inheritance in the co-optimization of soft robot morphology (via evolution) and controllers (via lifetime learning) within dynamic environments. The central claim is that Lamarckian inheritance underperforms Darwinian inheritance only when environmental changes are both conflicting and unpredictable; this is demonstrated via simulations employing Bayesian optimization and reinforcement learning. The authors further report that adding an environmental-change sensor restores Lamarckian benefits in conflicting settings by enabling agents to predict and generalize to new behaviors.

Significance. If the results hold under independent operationalization of the key variables, the work would help reconcile conflicting findings in the evolutionary robotics literature on the utility of Lamarckian inheritance. By isolating conflict and predictability as modulating factors, it supplies a testable framework for when lifetime learning with inheritance aids adaptation in changing conditions, with potential implications for algorithm design in real-world dynamic robotics.

major comments (2)
  1. [§3] §3 (Methods, Environment Construction): The labeling of environments as 'conflicting' or 'unpredictable' is described as arising from variations in task parameters (terrain or target shifts), yet the text indicates these labels are assigned based on observed performance gaps between Lamarckian and Darwinian runs. This risks circularity because the same simulation outcomes used to support the interaction effect are also used to define the independent variables. An a priori metric (e.g., controller distance or transition entropy computed before evolutionary trials) is required to substantiate the claim that underperformance occurs 'only when' both conditions hold.
  2. [§4] §4 (Results, Sensor Ablation): The restoration of Lamarckian benefits via an added sensor is presented as evidence that predictability can be engineered. However, without a pre-defined predictability measure independent of post-evolution performance, it remains unclear whether the sensor truly decouples the variables or simply alters the effective environment in a way that favors the reported outcome. Explicit quantification of predictability (e.g., mutual information between sensor readings and required controller changes) before and after sensor addition would strengthen the causal interpretation.
minor comments (2)
  1. [Abstract] The abstract and introduction could more explicitly state the two learning methods (Bayesian optimization and reinforcement learning) when first introducing the experimental setup, rather than deferring the detail.
  2. [Figures] Figure captions for the performance plots should include error bars or statistical test results to allow immediate assessment of the reported differences.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed and constructive review. The comments highlight important opportunities to improve the transparency of our variable definitions and causal claims. We address each major comment below and indicate the revisions planned for the next manuscript version.

read point-by-point responses

  1. Referee: §3 (Methods, Environment Construction): The labeling of environments as 'conflicting' or 'unpredictable' is described as arising from variations in task parameters (terrain or target shifts), yet the text indicates these labels are assigned based on observed performance gaps between Lamarckian and Darwinian runs. This risks circularity because the same simulation outcomes used to support the interaction effect are also used to define the independent variables. An a priori metric (e.g., controller distance or transition entropy computed before evolutionary trials) is required to substantiate the claim that underperformance occurs 'only when' both conditions hold.

    Authors: We agree that explicit a priori metrics are needed to avoid any appearance of circularity. Our environments are constructed by systematically varying task parameters (e.g., terrain friction coefficients or target displacement vectors) chosen in advance. 'Conflicting' is defined as cases where the Euclidean distance between optimal controller parameter vectors (obtained via separate Bayesian optimization or RL runs on each environment in isolation) exceeds a threshold, and cross-environment performance without re-learning drops below a set level; these quantities are computed before any evolutionary trials begin. 'Unpredictable' is defined via the absence of consistent temporal patterns in the parameter shifts. To address the concern directly, we will add a new subsection in Methods that reports these pre-computed metrics (controller distances and transition entropy) for each environment class and shows that the classification is fixed prior to the main experiments. This revision will make the independence from evolutionary outcomes explicit. revision: partial

  2. Referee: §4 (Results, Sensor Ablation): The restoration of Lamarckian benefits via an added sensor is presented as evidence that predictability can be engineered. However, without a pre-defined predictability measure independent of post-evolution performance, it remains unclear whether the sensor truly decouples the variables or simply alters the effective environment in a way that favors the reported outcome. Explicit quantification of predictability (e.g., mutual information between sensor readings and required controller changes) before and after sensor addition would strengthen the causal interpretation.

    Authors: We concur that an independent, quantitative measure of predictability would strengthen the interpretation. In the revised manuscript we will include a new analysis that computes the mutual information between the sensor observations and the required controller-parameter shifts across environment transitions. This calculation will be performed on the raw simulation trajectories generated before evolutionary optimization, both for the baseline agents and for agents equipped with the change-detecting sensor. The results will be reported in a new figure or table in the Sensor Ablation section, demonstrating that the added sensor measurably increases this mutual information and thereby supports the claim that it restores predictability rather than merely changing the environment in an ad-hoc manner. revision: yes

Circularity Check

0 steps flagged

No circularity: claims rest on direct simulation comparisons with a priori environment parameters

full rationale

The paper presents an empirical study using virtual soft robots and two learning methods (Bayesian optimization, reinforcement learning) to compare Lamarckian vs. Darwinian inheritance across environments varied by task parameters such as terrain or target shifts. The abstract and described methods define conflicting and unpredictable changes via these independent parameters before running trials, then measure performance outcomes. No equations, fitted parameters renamed as predictions, or load-bearing self-citations appear in the provided text that would reduce the central claim to a definitional tautology or post-hoc labeling. The derivation chain is self-contained against external simulation benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The work rests on standard domain assumptions from evolutionary robotics about inheritance mechanisms and simulation fidelity, without introducing new free parameters, axioms beyond the core Lamarckian transfer, or invented entities.

axioms (1)
  • domain assumption Lamarckian inheritance transfers learned controller parameters from parent to offspring in the evolutionary process
    This is the central methodological choice invoked throughout the abstract to contrast with Darwinian inheritance.

pith-pipeline@v0.9.0 · 5752 in / 1197 out tokens · 53179 ms · 2026-05-20T18:47:22.971146+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

48 extracted references · 48 canonical work pages

  1. [1]

    2006 , publisher=

    How the body shapes the way we think: a new view of intelligence , author=. 2006 , publisher=

    work page 2006

  2. [2]

    Artificial life , volume=

    Evolving 3D morphology and behavior by competition , author=. Artificial life , volume=

    work page

  3. [3]

    Nature , volume=

    Automatic design and manufacture of robotic lifeforms , author=. Nature , volume=. 2000 , publisher=

    work page 2000

  4. [4]

    Engineering Applications of Artificial Intelligence , volume=

    EDHMoR: Evolutionary designer of heterogeneous modular robots , author=. Engineering Applications of Artificial Intelligence , volume=. 2013 , publisher=

    work page 2013

  5. [5]

    Proceedings of the Artificial Life Conference 2016 (ALIFE XV) , author =

    On the. Proceedings of the Artificial Life Conference 2016 (ALIFE XV) , author =. 2016 , pages =

    work page 2016

  6. [6]

    Journal of The Royal Society Interface , volume=

    Scalable co-optimization of morphology and control in embodied machines , author=. Journal of The Royal Society Interface , volume=. 2018 , publisher=

    work page 2018

  7. [7]

    2016 , publisher=

    Evolutionary robotics , author=. 2016 , publisher=

    work page 2016

  8. [8]

    Frontiers in Robotics and AI , volume=

    Evolutionary robotics: what, why, and where to , author=. Frontiers in Robotics and AI , volume=. 2015 , publisher=

    work page 2015

  9. [9]

    Miras, Karine and De Carlo, Matteo and Akhatou, Sayfeddine and Eiben, A. E. , year =. Evolving-. Applications of

    work page

  10. [10]

    Frontiers in Robotics and AI , author =

    The. Frontiers in Robotics and AI , author =. 2022 , pages =

    work page 2022

  11. [11]

    Nature Communications , author =

    Embodied intelligence via learning and evolution , volume =. Nature Communications , author =. 2021 , pages =

    work page 2021

  12. [12]

    2025 IEEE Symposium on Computational Intelligence in Artificial Life and Cooperative Intelligent Systems (ALIFE-CIS) , pages=

    Generational Replacement and Learning for High-Performing and Diverse Populations in Evolvable Robots , author=. 2025 IEEE Symposium on Computational Intelligence in Artificial Life and Cooperative Intelligent Systems (ALIFE-CIS) , pages=. 2025 , organization=

    work page 2025

  13. [13]

    European conference on artificial life (ECAL-2013) , pages=

    The triangle of life: Evolving robots in real-time and real-space , author=. European conference on artificial life (ECAL-2013) , pages=

    work page 2013

  14. [14]

    Scientific Reports , author =

    Enhancing robot evolution through. Scientific Reports , author =. 2023 , pages =

    work page 2023

  15. [15]

    Frontiers in Robotics and AI , author =

    Lamarckian. Frontiers in Robotics and AI , author =

    work page

  16. [16]

    Proceedings of the Genetic and Evolutionary Computation Conference , pages=

    Lamarckian Co-design of Soft Robots via Transfer Learning , author=. Proceedings of the Genetic and Evolutionary Computation Conference , pages=

    work page

  17. [17]

    Artificial Life Conference Proceedings 32 , pages=

    Mechanisms of social learning in evolved artificial life , author=. Artificial Life Conference Proceedings 32 , pages=. 2020 , organization=

    work page 2020

  18. [18]

    Proceedings of the 2015 annual conference on genetic and evolutionary computation , pages=

    Three-fold adaptivity in groups of robots: the effect of social learning , author=. Proceedings of the 2015 annual conference on genetic and evolutionary computation , pages=

    work page 2015

  19. [19]

    Anthropological Science , volume=

    Individual versus social learning: evolutionary analysis in a fluctuating environment , author=. Anthropological Science , volume=. 1996 , publisher=

    work page 1996

  20. [20]

    , author=

    Balancing the Costs and Benefits of Learning Ability. , author=. ECAL , pages=

    work page

  21. [21]

    European Conference on Artificial Life , pages=

    On the adaptive disadvantage of Lamarckianism in rapidly changing environments , author=. European Conference on Artificial Life , pages=. 2007 , organization=

    work page 2007

  22. [22]

    Artificial Life , volume=

    Evolving learnable neural networks under changing environments with various rates of inheritance of acquired characters: comparison of Darwinian and Lamarckian evolution , author=. Artificial Life , volume=. 1999 , publisher=

    work page 1999

  23. [23]

    IEEE Transactions on Evolutionary Computation , year=

    Lamarckian Inheritance Improves Robot Evolution in Dynamic Environments , author=. IEEE Transactions on Evolutionary Computation , year=

    work page

  24. [24]

    Advances in Neural Information Processing Systems , volume=

    Evolution gym: A large-scale benchmark for evolving soft robots , author=. Advances in Neural Information Processing Systems , volume=

    work page

  25. [25]

    Proceedings of the Genetic and Evolutionary Computation Conference , pages=

    Modular controllers facilitate the co-optimization of morphology and control in soft robots , author=. Proceedings of the Genetic and Evolutionary Computation Conference , pages=

    work page

  26. [26]

    International Conference on the Applications of Evolutionary Computation (Part of EvoStar) , pages=

    Out of time: On the constrains that evolution in hardware faces when evolving modular robots , author=. International Conference on the Applications of Evolutionary Computation (Part of EvoStar) , pages=. 2022 , organization=

    work page 2022

  27. [27]

    Journal of the Royal statistical society: series B (Methodological) , volume=

    Controlling the false discovery rate: a practical and powerful approach to multiple testing , author=. Journal of the Royal statistical society: series B (Methodological) , volume=. 1995 , publisher=

    work page 1995

  28. [28]

    The annals of mathematical statistics , pages=

    On a test of whether one of two random variables is stochastically larger than the other , author=. The annals of mathematical statistics , pages=. 1947 , publisher=

    work page 1947

  29. [29]

    Proceedings of the Genetic and Evolutionary Computation Conference , pages=

    How the morphology encoding influences the learning ability in body-brain co-optimization , author=. Proceedings of the Genetic and Evolutionary Computation Conference , pages=

    work page

  30. [30]

    2011 , publisher=

    Hiller, Jonathan and Lipson, Hod , journal=. 2011 , publisher=

    work page 2011

  31. [31]

    2023 , publisher=

    Legrand, Julie and Terryn, Seppe and Roels, Ellen and Vanderborght, Bram , journal=. 2023 , publisher=

    work page 2023

  32. [32]

    2024 , organization=

    Mertan, Alican and Cheney, Nick , booktitle=. 2024 , organization=

    work page 2024

  33. [33]

    Proceedings of the 2020 Genetic and Evolutionary Computation Conference , pages=

    Evolution of distributed neural controllers for voxel-based soft robots , author=. Proceedings of the 2020 Genetic and Evolutionary Computation Conference , pages=

    work page 2020

  34. [34]

    European Conference on Genetic Programming (Part of EvoStar) , pages=

    Grammar-Based Evolution of Polyominoes , author=. European Conference on Genetic Programming (Part of EvoStar) , pages=. 2024 , organization=

    work page 2024

  35. [35]

    ACM SIGEVOlution , volume=

    Unshackling evolution: evolving soft robots with multiple materials and a powerful generative encoding , author=. ACM SIGEVOlution , volume=. 2014 , publisher=

    work page 2014

  36. [36]

    IEEE Transactions on Evolutionary Computation , year=

    Enhancing Adaptability in Embodied Agents: A Multi-Quality-Diversity Approach , author=. IEEE Transactions on Evolutionary Computation , year=

    work page

  37. [37]

    ACM Transactions on Evolutionary Learning , volume=

    Factors impacting diversity and effectiveness of evolved modular robots , author=. ACM Transactions on Evolutionary Learning , volume=. 2023 , publisher=

    work page 2023

  38. [38]

    IEEE Transactions on Evolutionary Computation , volume=

    Quality and diversity optimization: A unifying modular framework , author=. IEEE Transactions on Evolutionary Computation , volume=. 2017 , publisher=

    work page 2017

  39. [39]

    arXiv preprint arXiv:2508.17464 , year=

    Evolutionary Brain-Body Co-Optimization Consistently Fails to Select for Morphological Potential , author=. arXiv preprint arXiv:2508.17464 , year=

    work page arXiv

  40. [40]

    Proceedings of the Genetic and Evolutionary Computation Conference , pages=

    Controller Distillation Reduces Fragile Brain-Body Co-Adaptation and Enables Migrations in MAP-Elites , author=. Proceedings of the Genetic and Evolutionary Computation Conference , pages=

    work page

  41. [41]

    2019 , eprint=

    Continuous control with deep reinforcement learning , author=. 2019 , eprint=

    work page 2019

  42. [42]

    The Quarterly review of biology , volume=

    Transgenerational epigenetic inheritance: prevalence, mechanisms, and implications for the study of heredity and evolution , author=. The Quarterly review of biology , volume=. 2009 , publisher=

    work page 2009

  43. [43]

    Annual Review of Ecology, Evolution, and Systematics , volume=

    Nongenetic inheritance and its evolutionary implications , author=. Annual Review of Ecology, Evolution, and Systematics , volume=. 2009 , publisher=

    work page 2009

  44. [44]

    Evolution , volume=

    Genotype-environment interaction and the evolution of phenotypic plasticity , author=. Evolution , volume=. 1985 , publisher=

    work page 1985

  45. [45]

    ACM Transactions on mathematical software (TOMS) , volume=

    Algorithm 778: L-BFGS-B: Fortran subroutines for large-scale bound-constrained optimization , author=. ACM Transactions on mathematical software (TOMS) , volume=. 1997 , publisher=

    work page 1997

  46. [46]

    and Roijers, Diederik M

    Lan, Gongjin and De Carlo, Matteo and Van Diggelen, Fuda and Tomczak, Jakub M. and Roijers, Diederik M. and Eiben, A.E. , month = nov, year =

    work page

  47. [47]

    2024 , publisher=

    van Diggelen, Fuda and Ferrante, Eliseo and Eiben, AE , journal=. 2024 , publisher=

    work page 2024

  48. [48]

    Proceedings of the Genetic and Evolutionary Computation Conference (GECCO 2026) , year =

    de Bruin, K Ege and Glette, Kyrre and Ellefsen, Kai Olav and Nadizar, Giorgia and Medvet, Eric , title =. Proceedings of the Genetic and Evolutionary Computation Conference (GECCO 2026) , year =

    work page 2026