Video generation models are increasingly seen as a promising foundation for embodied world models, but their usefulness hinges on whether the rollouts they produce are physically plausible and aligned with the given task instructions. According to the abstract for RoboJudge, a new framework from arXiv, multimodal large language models (MLLMs) are being explored as judges for exactly this kind of evaluation.
The core idea is to use MLLMs to assess generated videos, rather than relying solely on human evaluation or hand-crafted metrics. This could provide a scalable way to check whether a generated rollout obeys the laws of physics and follows the user's intent, both of which are critical for real-world embodied applications.
The abstract is brief and does not include experimental details or performance numbers, so the specific methodology and results remain to be seen in the full paper. Still, the direction highlights a growing interest in using LLM-based judges for generative model evaluation, now extended to the embodied video domain.