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

In robotics, "brain-body co-optimization" refers to the simultaneous design of a robot's physical structure (morphology) and its control software. Despite its intuitive appeal, the process is notoriously difficult due to a "chicken and egg" problem: a superior body requires a specialized controller to demonstrate its value, while a sophisticated controller is limited by the physical constraints of the body. Historically, it has been unclear whether failures in co-optimization stem from the vastness of the search space, the ruggedness of the fitness landscape, or fundamental flaws in evolutionary dynamics. This study addresses these questions by exhaustively mapping a morphology-fitness landscape to observe exactly how algorithms navigate it.

Methods and Dataset

The researchers utilized Evogym, a 2D voxel-based soft robot simulator, to define a design space of 1,305,840 distinct morphologies within a 3x3 grid. Each morphology was composed of four possible materials: soft/rigid passive voxels or horizontal/vertical active voxels.

To establish a ground-truth landscape, a neural network controller was trained for every possible morphology for 300 generations. This exhaustive mapping, inspired by benchmarks like NAS-Bench-101, provides a complete picture of the "true potential" of every design in the space. The study then evaluated two prominent evolutionary algorithms—Age-Fitness Pareto Optimization (AFPO) and MAP-Elites—on their ability to find these high-potential designs when evolving brains and bodies simultaneously.

The Challenge of Fragile Co-adaptation

A key finding is the phenomenon of fragile co-adaptation. When a morphology mutates, the existing controller typically becomes ineffective, leading to an immediate drop in fitness. Because evolutionary algorithms prioritize current performance, they often discard these "unfit" mutations before the controller has time to adapt to the new body shape.

Correlation between current fitness and true potential over 300 generations of controller training.

As shown in the figure above, the correlation between a morphology's fitness during early training and its eventual "true potential" is initially very low. It takes nearly the full 300 generations for the correlation to approach 1.0, particularly for the top-performing designs. This lag means that algorithms consistently "undervalue" new, potentially superior body shapes, favoring mediocre but well-trained "veteran" designs instead.

Key Results

The exhaustive map revealed that while the landscape is rugged, ruggedness is not the primary obstacle. Instead, algorithms frequently get stuck at morphologies that are not even local maxima—meaning a better design was just one mutation away, but the algorithm couldn't "see" it.

Fitness distributions for AFPO and MAP-Elites compared to the near-optimal threshold (dashed line).

In 100 independent trials, AFPO reached a near-optimal morphology (top 15%) only 7% of the time, while MAP-Elites succeeded in 13% of runs. As illustrated in the comparison above, the majority of evolved solutions fall significantly below the potential threshold (the dashed line). Strikingly, when the same algorithms were run using the "true potential" (pre-calculated fitness) instead of simultaneous training, they found optimal solutions 100% of the time. This confirms that the search space is navigable; the failure lies entirely in the co-optimization process.

Conclusion and Future Work

The failure of evolutionary brain-body co-optimization is largely driven by the inability of algorithms to track the fitness gradient in the face of fragile co-adaptation. Future work must focus on alleviating this "negative transfer" between body mutations and controllers. Potential solutions include developing "universal controllers," implementing "grace periods" for new morphologies to learn before selection, or utilizing differentiable simulations to bypass the fitness valleys created by the estimation gap.

Summary updated

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.
BibTeX
@misc{mertan2025evolutionary,
  author = {Alican Mertan and Nick Cheney},
  title = {{Evolutionary Brain-Body Co-Optimization Consistently Fails to Select for Morphological Potential}},
  year = {2025},
  eprint = {2508.17464},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2508.17464},
}
Figure