Evolutionary Brain-Body Co-Optimization Consistently Fails to Select for Morphological Potential

Alican Mertan, Nick Cheney · 2025

Read Evolutionary Brain-Body Co-Optimization Consistently Fails to Select for Morphological Potential by Alican Mertan, Nick Cheney (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2508-17464.txt.

Introduction

Robots evolved in simulation often pair a body with a controller, and breeders hope the search favours bodies with room to grow. Alican Mertan and Nick Cheney tested that hope by mapping the whole landscape of one design space: more than 1.3 million soft voxel robots, each trained with its own controller. With the full map in hand, they could watch evolutionary searches wander it.

The searches kept stalling. New body mutations were scored before their controllers matured, so promising shapes were culled early, sometimes one step from a better neighbour. On the bright side, changing bodies mid-search opened goals that fixed-body training never reached. The authors say the field's common failure has a concrete cause, and their map gives future algorithms a fair test bed.

Catalog record

CREATORAlican Mertan, Nick Cheney
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE37 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2508-17464

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Evolutionary Brain-Body Co-Optimization Consistently Fails to Select for Morphological Potential Alican Mertan, Nick Cheney Abstract Brain-body co-optimization remains a challenging problem, despite increasing interest from the community in recent years. To understand and overcome the challenges, we propose exhaustively mapping a morphology-fitness landscape to study it. To this end, we train controllers for each feasible morphology in a design space of 1,305,840 distinct morphologies, constrained by a computational budget. First, we show that this design space constitutes a good model for studying the brain-body co-optimization problem, and our attempt to exhaustively map it roughly captures the landscape. We then proceed to analyze how evolutionary brain-body co-optimization algorithms work in this design space. The complete knowledge of the morphology-fitness landscape facilitates a better understanding of the results of evolutionary brain-body co-optimization algorithms and how they unfold over evolutionary time in the morphology space. This investigation shows that the experimented algorithms cannot consistently find near-optimal solutions. The search, at times, gets stuck on morphologies that are sometimes one mutation away from better morphologies, and the algorithms cannot efficiently track the fitness gradient in the morphology-fitness landscape. We provide evidence that experimented algorithms regularly undervalue the fitness of individuals with newly mutated bodies and, as a result, eliminate promising morphologies throughout evolution. Our work provides the most concrete demonstration of the challenges of evolutionary brain-body co-optimization. Our findings ground the trends in the literature and provide valuable insights for future work. Introduction Brain-body co-optimization refers to the process of optimizing both the control and design of a robot for a given task. While the origins of brain-body co-optimization go back more than 30 years ago ( Sims, 1994 ) , there has been a recent increase in research interest ( Liu et al., 2025 ; Wang, 2025 ) . Especially with the advent of soft robots consisting of materials capable of highly complex and unintuitive behavior ( Hiller and Lipson, 2012 ) , the need for automated design and control has increased greatly. Moreover, studies of embodied intelligence have shown how specialized designs for a given task and environment can improve the efficacy and efficiency of robot designs, find more robust solutions, or find ingenious designs that would be difficult to design by hand ( Pfeifer and Bongard, 2007 ; Bhatia et al., 2021 ) . Brain-body co-optimization, however, still remains a challenging problem despite great interest from the research community. In this work, our objective is to improve our understanding of the problem and facilitate further research in this domain. Previous literature offers a general approach to such a problem, where low-level features of fitness landscapes are estimated to characterize them ( Mersmann et al., 2010 ; Mersmann et al., 2011 ) . Here, to gain a deeper understanding of a specific optimization problem, we propose a different, more computationally demanding approach. Inspired by the success of the NAS-Bench-101 public architecture dataset and its derivatives in fostering more research and benefiting the neural architecture search community ( Ying et al., 2019 ; Dong and Yang, 2020 ) , we set out to create a similar dataset where we measure the true fitness of each morphology in a given design space, building an exhaustive morphology-fitness landscape. To map the morphology-fitness landscape exhaustively, we train controllers for each feasible morphology in a design space of 1,305,840 distinct morphologies, constrained by a computational budget. We use the best performance achieved during the training as an estimate of their best possible performance. We investigate the validity and significance of the data and conclude that it provides a good test bed to study the brain-body co-optimization problem. We believe that these data have the potential to improve our understanding by providing complete access to the search space. To make an initial attempt to use these data to gain insights into the brain-body co-optimization problem, we investigate how two commonly used evolutionary optimization algorithms, namely age-fitness Pareto optimization ( Schmidt and Lipson, 2010 ) and MAP-Elites ( Mouret and Clune, 2015 ) , work on this morphology-fitness landscape. Having access to the complete morphology-fitness landscape allows us to demonstrate that the evolutionary optimization algorithms tested often fail to find near-optimal solutions. Most strikingly, we show that both algorithms, at times, get stuck in morphologies that are one mutation away from a morphology with better performance. We show that both algorithms frequently eliminate morphologies with better performance due to their inability to estimate true potential. Although previous research has provided evidence suggesting this phenomenon – fragile co-adaptation ( Joachimczak et al., 2016 ; Cheney et al., 2016 ; Mertan and Cheney, 2024 ) , having access to the full morphology-fitness landscape enables us to clearly illustrate fragile co-adaptation and its implications. Lastly, we review the existing literature in light of our findings, suggesting potential future work directions, both to make use of the provided data and to overcome challenges of brain-body co-optimization. In summary, our contributions are as follows. • We provide a dataset of a complete morphology-fitness landscape -- estimated true fitness values of all feasible voxel-based soft robots in a design space. We provide evidence that the mapped morphology-fitness landscape constitutes a good model to study the co-optimization problem. We hope that this data facilitates further research in this domain 1 1 1 The data and code are available at our repository. (Fig. 1 , 2 , 3 ). • We investigate how two different evolutionary brain-body co-optimization algorithms, namely age-fitness Pareto optimization ( Schmidt and Lipson, 2010 ) and MAP-Elites ( Mouret and Clune, 2015 ) , work by using complete knowledge of the morphology-fitness landscape, providing the most concrete evidence to date (Section Brain-body co-optimization ). • We show that ranking high-performing individuals correctly requires extensive controller training and hypothesize that this need, combined with fragile co-adaptation, hinders the optimization of morphologies in evolutionary brain-body co-optimization by creating dynamics leading to first-mover advantages (Fig. 3 , 7 ). • We show that search, at times, stagnates at points in the morphology-fitness landscape that are not even a local maximum. We provide evidence to show that evolutionary brain-body co-optimization algorithms consistently undervalue the fitness of offspring with newly mutated bodies, eliminating promising morphologies (Fig. 4 , 5 , 6 , 7 ). • We use our findings to ground trends in the literature and identify promising research directions (Section Discussion ). Methodology Simulation Engine We use Evogym ( Bhatia et al., 2021 ) as a simulation engine, which is a 2D voxel-based soft body simulator. Each voxel is modeled by 4 point masses and 6 springs connecting them. Brain-body co-optimization The problem we are interested in analyzing is the optimization of control and morphology, that is, finding the best-performing morphology-controller pair for a given environment and task. Task and fitness The task we consider in this work is a locomotion task. The environment consists of a flat surface and a target position that robots are tasked to reach. Modified version of the fitness described in ( Bhatia et al., 2021 ) is used, where an additional small negative penalty ( =-0.05 ) is applied each time step before the robot reaches the target to facilitate the selection of robots that move towards the goal position as

Read the complete text