Developmental encodings are often motivated by the expectation that a structured genotype-to-phenotype process can improve evolvability, robustness, and adaptation. Yet an encoding that expands phenotypic variation may simultaneously make useful parental structure harder to preserve under mutation and recombination. We test this trade-off in a controlled evolutionary-robotics benchmark comparing three matched representations: direct encoding, a static generative encoding, and a temporal zygotic developmental encoding. All three conditions use 34 genome parameters, 34 adult-controller parameters, identical initial genome matrices, the same optimizer, the same mutation and crossover operators, the same environments, and the same number of fitness evaluations. Across 30 paired evolutionary runs per representation, developmental encoding does not reliably exceed direct encoding in final training or out-of-distribution fitness. It does, however, substantially increase final population phenotypic diversity and mutant-reachable phenotype diversity. The same representation exhibits lower mutation viability, lower genotype--phenotype locality, sharply reduced crossover offspring fitness and viability, and increased recombination novelty. Continued developmental dynamics also attenuate damage applied early in development relative to otherwise identical late damage. These results identify a representation-level trade-off: development can enlarge the neighborhood of reachable phenotypes and provide within-development damage attenuation while degrading the local inheritance of already-adapted structure. We argue that developmental encodings should therefore be evaluated not by performance alone, but by a joint geometry of generativity, locality, robustness, and inheritance compatibility.
Figures & tables
Figure 1: Conceptual contrast between direct and zygotic developmental encoding. Direct encoding evolves the finished controller, whereas developmental encoding evolves an inherited code seed that is transformed through a starting state and temporal self-organization before the adult controller is produced. The static generative baseline used in the experiments is omitted here for visual clarity. Genome dimension, mutation, recombination, selection, evaluation budget, and adult-controller dimension are matched across the benchmark; only the genotype-to-phenotype map differs.
Figure 2: Evolutionary learning curves for the three representation classes across matched runs. Development changes the trajectory and variability of search but does not yield a reliable final-fitness advantage over direct encoding in the paired analysis.
Figure 3: Local mutation analysis. Top left: mutation viability, Eviable=P(F(g′)>0) . Top right: beneficial-mutation probability, Ebeneficial=P(F(g′)>F(g)) . Bottom left: distribution of phenotype distance dP after one mutation. Bottom right: genotype–phenotype locality, corr(dG,dP) . In the three summary panels, small points denote individual runs and large markers with error bars denote the median and bootstrap 95% confidence interval.
Figure 4: Recombination outcomes. Top left: relative crossover fitness, Rcross=Fchild−Fparents ; the dotted horizontal line marks mid-parent parity. Top right: fraction of viable offspring. Bottom: offspring fitness versus mid-parent fitness for all recombination trials; the dashed diagonal denotes equal offspring and mid-parent fitness. The developmental representation shows substantially lower crossover fitness and offspring viability.
Figure 5: Perturbation and developmental-damage analyses. Damage introduced early during the developmental trajectory is strongly attenuated relative to late damage, while adult module failure is also somewhat less costly under the developmental representation.
Can evolutionary dynamics characteristic of biological development arise without a designer-specified fitness function? We present Genesis, a platform in which agents inhabit a Gray-Scott reaction-diffusion substrate and evolve under physical constraints alone. Three successive experimental cycles, each testing one falsifiable hypothesis, show: (1) constraint-driven selection sustains evolutionary activity after complete fitness removal but reaches a hard phenotypic complexity ceiling; (2) agent-mediated niche construction via chemical secretion is real but causally insufficient to break that ceiling; and (3) replacing the fixed-alphabet genome with a Compositional Pattern Producing Network (CPPN) indirect encoding, protected by NEAT-style speciation, produces the first evidence of progressive structural complexification in a fitness-free system. Null results are treated as precise, informative answers rather than failures, yielding reusable diagnostic tools and a sham-control protocol applicable to any open-ended evolution evaluation pipeline.
Lamarckian inheritance has been shown to be a powerful accelerator in systems where the joint evolution of robot morphologies and controllers is enhanced with individual learning. Its defining advantage lies in the offspring inheriting controllers learned by their parents. The efficacy of this option, however, relies on morphological similarity between parent and offspring. In this study, we examine how Lamarckian inheritance performs when the search process is driven toward high morphological variance, potentially straining the requirement for parent-offspring similarity. Using a system of modular robots that can evolve and learn to solve a locomotion task, we compare Darwinian and Lamarckian evolution to determine how they respond to shifting from pure task-based selection to a multi-objective pressure that also rewards morphological novelty. Our results confirm that Lamarckian evolution outperforms Darwinian evolution when optimizing task-performance alone. However, introducing selection pressure for morphological diversity causes a substantial performance drop, which is much greater in the Lamarckian system. Further analyses show that promoting diversity reduces parent-offspring similarity, which in turn reduces the benefits of inheriting controllers learned by parents. These results reveal the limits of Lamarckian evolution by exposing a fundamental trade-off between inheritance-based exploitation and diversity-driven exploration.
Jed R Muff, Karine Miras, A. E. Eiben
Department of Computer Science Vrije Universiteit Amsterdam, The Netherlands
In this paper, we introduce a model of evolution and learning in robots that co-optimizes a distribution of latent design vectors (genotypes) and a mixture of control experts (neural modules), which are gated by the latent coordinates of each decoded design (phenotype). This provides a scalable alternative to co-design algorithms that either train an individual policy for every robot, which is inefficient, or a monolithic universal controller for all robots, which results in overly conservative structures and behaviors. Our approach lies somewhere between these two extremes, preserving ancestral knowledge in a unified yet modular framework in which different body plans activate and deactivate different combinations of learned sensorimotor circuits for goal-directed behavior. This allows one part of the controller to be overhauled to better suit new species of designs as they emerge without disrupting the hard-earned knowledge contained within other expert modules. It also allows pretrained expert policies to be directly plugged into the mixture, which can steer evolution into otherwise unexplored areas of latent space containing desired morphological traits. We refer to this process as "evo by demo" and explore how it may be used to guide freeform evolution toward canonical structures defined by the pretrained model. Videos and code can be found at: https://eco-moe.github.io.