Abstract
We introduce LLM-Morph, an automated approach for generating and optimizing robot designs using large language models (LLMs) and evolutionary algorithms.
In this framework, each stage of the robot's morphology optimization process is guided by Large Language Models (LLMs), replacing the traditional neural network-based design approach from the initial design phase onward. By integrating automatic prompt design, our LLM-Morph framework continuously enhances control effectiveness by refining robot designs and improving overall performance.
Our experimental results demonstrate that LLM-Morph effectively generates innovative robot designs and optimizes their performance enhancements. Additionally, LLM-Morph can further enhance performance simply by fine-tuning the morphology. Our approach demonstrates the potential of using LLMs for data-driven and modular robot design, providing a promising methodology that can be extended to other domains with similar design frameworks.
The LLM designs a robot based on the user’s requirements, including the intended application and preferences. The design is then converted into an XML file using a mini-program.
Experiments and Results
We ran five experiments, each focusing on a specific aspect:
Speed, Distance, and Energy aim to improve the robot's performance, while Change Size and Change Location focus on optimizing fitness.
Speed
12
13
20
21
Distance
60
117
150
163
Energy
-3.2
1
1.6
8.5
Fitness Change Size
6.5
17
18
21
Fitness Change Location
6.5
30
43
46
BibTeX
@article{
}