Research Analyzer
← Back ICRA 2026

RoboMorph: Evolving Robot Morphology Using Large Language Models

Kevin Qiu, Władysław Pałucki, Krzysztof Ciebiera, Paweł Fijałkowski, Marek Cygan, Łukasz Kuciński

PDF

AI summary

Key figure (auto-extracted from paper)
LLMs effectively serve as generative mutation operators in evolutionary loops to automatically discover high-performing, terrain-specialized robot morphologies.
Robot morphology Large language models Evolutionary algorithms Automated design Reinforcement learning Robotics

Problem

Automating robot morphology design is traditionally time-consuming, computationally expensive, and limited by fixed heuristics that struggle to explore vast design spaces.

Approach

The framework uses large language models to propose structurally valid robot designs, which are evaluated via reinforcement learning in simulation and iteratively refined through an evolutionary feedback loop.

Key results

  • Discovers diverse, terrain-specialized morphologies across four environments
  • Matches or outperforms the Robogrammar baseline under RL and MPC control
  • Demonstrates LLMs as efficient generative mutation operators for design space exploration
  • Validates physical plausibility by retaining top designs under full collision physics

Why it matters

Enables robotics researchers to rapidly automate and scale morphology discovery, reducing development time and uncovering novel designs beyond human intuition.

Abstract

We introduce RoboMorph, an automated ap- proach for generating and optimizing modular robot designs using large language models (LLMs) and evolutionary al- gorithms. Each robot design is represented by a structured grammar, and we use LLMs to efficiently explore this design space. Traditionally, such exploration is time-consuming and computationally intensive. Using a best-shot prompting strat- egy combined with reinforcement learning (RL)-based control evaluation, RoboMorph iteratively refines robot designs within an evolutionary feedback loop. Across four terrain types, RoboMorph discovers diverse, terrain-specialized morpholo- gies, including wheeled quadrupeds and hexapods, that match or outperform designs produced by Robogrammar’s graph- search method. These results demonstrate that LLMs, when coupled with evolutionary selection, can serve as effective gen- erative operators for automated robot design. Our project page and code are available at https://robomorph.github.io.

Index terms

Methods and Tools for Robot System Design Evolutionary Robotics Cellular and Modular Robots

Related papers