RoboMorph: Evolving Robot Morphology using Large Language Models

Kevin Qiu 1,2 Władysław Pałucki 1 Krzysztof Ciebiera 1 Paweł Fijałkowski 1 Marek Cygan 1,3 Łukasz Kuciński 1,2,4
1University of Warsaw 2IDEAS NCBR 3Nomagic 4Polish Academy of Sciences
ICRA 2026
Paper Code arXiv
Overview of the RoboMorph framework
Overview of the RoboMorph framework. The LLM proposes designs within an evolutionary loop, where each candidate is compiled, trained via RL, evaluated, and fed back into the population.

TL;DR: We use LLMs as mutation operators in an evolutionary loop to automatically design terrain-specialized robots.

Abstract

We introduce RoboMorph, an automated approach for generating and optimizing modular robot designs using large language models (LLMs) and evolutionary algorithms. Each robot design is represented by a structured grammar, and we use LLMs to efficiently explore this design space. Using a best-shot prompting strategy 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 morphologies, 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 generative operators for automated robot design.

Top Designs

Ridged
Flat
Frozen
Beams
Top robot designs evolved by RoboMorph for four terrains: ridged, flat, frozen, and beams
Quadruped with high foot clearance
Wheeled quadruped with rolling contact
Hexapod with wide, stable stance
Compact low-profile crawler

Results

Average maximum fitness across evolutionary steps for all four terrains
Average maximum fitness (95% CI, 8 seeds) at each evolutionary step. Fitness improves consistently across generations.
Evolution trajectory of the best-performing seed for each terrain
Evolution of the best-performing seed for each terrain, showing morphologies discovered at key generations.

Comparison with Robogrammar

RoboMorph matches or outperforms Robogrammar across all terrains under both RL and MPC control.

RoboMorph vs Robogrammar under RL control
RL control.
RoboMorph vs Robogrammar under MPC control
MPC control.

Citation

@inproceedings{qiu2026robomorph,
  title={RoboMorph: Evolving Robot Morphology using Large Language Models},
  author={Qiu, Kevin and Pa{\l}ucki, W{\l}adys{\l}aw and Ciebiera, Krzysztof and Fija{\l}kowski, Pawe{\l} and Cygan, Marek and Kuci{\'n}ski, {\L}ukasz},
  booktitle={IEEE International Conference on Robotics and Automation (ICRA)},
  year={2026}
}