Efficient Robot Design with Multi-Objective Black-Box Optimization and Large Language Models

TL;DR

The paper tackles the challenge of efficiently designing robot bodies under multiple objectives by marrying multi-objective black-box optimization with large language model (LLM)–based sampling. It defines design parameters (base position $\bm{O}$, joint types $\bm{M}$, link lengths $\bm{L}$) and objectives ($E^{pos}$ and $E^{torque}$) evaluated via inverse kinematics, and uses the hypervolume $E_{hv}$ to assess Pareto coverage. The core contribution is a hybrid framework where LLM-based sampling (with problem settings and feedback) guides design exploration in parallel with Tree-Structured Parzen Estimator-based MO optimization, showing improved exploration and more stable Pareto fronts, particularly when using prompts with specific parameter analysis ($\text{LLM+}$). Experimental results across three target operation points demonstrate that modest interleaving of LLM proposals ($N_{step}=10$) yields the best performance, enabling symmetric workspaces and designs that are hard to discover with BBO alone. The work highlights both the potential and limitations of integrating LLMs into robot design optimization, pointing to future work on handling more complex constraints and personalized task-driven designs.

Abstract

Various methods for robot design optimization have been developed so far. These methods are diverse, ranging from numerical optimization to black-box optimization. While numerical optimization is fast, it is not suitable for cases involving complex structures or discrete values, leading to frequent use of black-box optimization instead. However, black-box optimization suffers from low sampling efficiency and takes considerable sampling iterations to obtain good solutions. In this study, we propose a method to enhance the efficiency of robot body design based on black-box optimization by utilizing large language models (LLMs). In parallel with the sampling process based on black-box optimization, sampling is performed using LLMs, which are provided with problem settings and extensive feedback. We demonstrate that this method enables more efficient exploration of design solutions and discuss its characteristics and limitations.

Figures (12)

• Figure 1: The concept of this study: high efficiency robot design with multi-objective black-box optimization and LLM-based sampling.
• Figure 2: The overview of the proposed robot design optimization. We set the robot design parameters, automatically generate the robot URDF, compute the objective functions, and perform multi-objective black-box optimization and LLM-based sampling.
• Figure 3: The robot design parameters handled in this study. The parameters include the base link coordinate $\bm{O}$, the joint types $\bm{M}$, and the link lengths $\bm{L}$ of the robot.
• Figure 4: Hypervolume for the evaluation of multi-objective optimization.
• Figure 5: The overview of the prompt for LLM-based sampling conducted in this study.