Debate2Create: Robot Co-design via Large Language Model Debates

TL;DR

This work tackles joint morphology and control co-design in robotics, framed as a high-dimensional optimization over morphology $m \in \mathcal{M}$ and reward function $r \in \mathcal{R}$ to maximize $S(m,\pi^*(m,r))$. It introduces Debate2Create (D2C), a multi-agent LLM debate framework where a design agent edits morphology and a control agent crafts a morphology-conditioned reward, with physics-grounded evaluation guiding the debate. A key contribution is formalizing the objective $\max_{m \in \mathcal{M}, r \in \mathcal{R}} S(m,\pi^*(m,r))$ and leveraging simulator feedback and a hall-of-fame to steer iteration and avoid reward hacking. On the Brax Ant locomotion task, D2C discovers diverse, higher-performing morphologies and achieves substantial gains over the baseline (top score $6421.67$ vs $3715.42$), demonstrating the effectiveness of structured LLM debates for automated robot co-design.

Abstract

Automating the co-design of a robot's morphology and control is a long-standing challenge due to the vast design space and the tight coupling between body and behavior. We introduce Debate2Create (D2C), a framework in which large language model (LLM) agents engage in a structured dialectical debate to jointly optimize a robot's design and its reward function. In each round, a design agent proposes targeted morphological modifications, and a control agent devises a reward function tailored to exploit the new design. A panel of pluralistic judges then evaluates the design-control pair in simulation and provides feedback that guides the next round of debate. Through iterative debates, the agents progressively refine their proposals, producing increasingly effective robot designs. Notably, D2C yields diverse and specialized morphologies despite no explicit diversity objective. On a quadruped locomotion benchmark, D2C discovers designs that travel 73% farther than the default, demonstrating that structured LLM-based debate can serve as a powerful mechanism for emergent robot co-design. Our results suggest that multi-agent debate, when coupled with physics-grounded feedback, is a promising new paradigm for automated robot design.

Figures (2)

• Figure 1: Overview of the Debate2Create framework. (A) A dialectical debate between the design agent () and control agent () to propose and critique morphology--reward hypotheses. (B) A physics simulator evaluates each proposed design--control pair, and a panel of pluralistic judges () reasons over the resulting performance metrics to provide feedback. (C) A hall-of-fame archive stores the best design--control pairs from each round to inform subsequent debates. Takeaway: Multi-agent debate enables discovery of novel robot morphologies and control strategies that single-agent methods would miss.
• Figure 2: Performance of D2C over debate rounds, showing forward-distance score $S$ for thesis vs. synthesis designs at each round. Each round, the thesis and synthesis morphologies are evaluated under the control agent's proposed reward. Error bars denote 95% confidence intervals across multiple reward candidates per design. Takeaway: Synthesis consistently outperforms thesis, indicating that the dialectical debate (thesis--antithesis--synthesis) yields progressively better designs.