A new preprint on arXiv (2610.27865) introduces RoboColiseum, an evaluation platform for studying frontier multimodal large language models as robot controllers. The work responds to a growing body of research on using LLMs for robot control, which the authors say has not yet produced a clear picture of how these models perform across different models, tasks, and execution settings.
The proposed evaluation is designed to be cross-model, meaning it places multiple LLMs in comparable robot-control scenarios rather than testing a single model in isolation. The abstract frames this as a necessary step toward understanding the embodied capabilities of LLMs—what they can actually do when tasked with controlling a robot.
The announcement does not include detailed results, so the paper's findings are not yet available from the abstract alone. The contribution at this stage is the evaluation framework itself and the goal of systematically characterizing frontier LLMs in robotics.