Impact of LLM-Assisted Coding in Creativity and Robustness of Robot Controllers

The use of Large Language Models (LLMs) in work environments has recently started to gain attention in the research community, with many works reporting increased productivity and others reporting homogenization of the output products. Nevertheless, their use as coding assistants in robotics has been mostly overlooked, especially from the point of view of their effects compared to not-assisted programming. We claim that peculiar characteristics of robotics programming deserve special attentions, such as the robustness of the produced solution. Here we analyze the effects of using LLMs as coding assistants in robotics. We analyze their effects on the performance, in a pseudo-reality gap, and on the similarity of the produced controllers. We also briefly discuss the feedback of some participants of the experiment. The results suggest that the codes produced with the assistance of LLMs are less robust to unseen conditions, and overall more homogeneous. Additionally, we report a shorter development time when using LLMs, but a poorer coding experience.