RoboGaze: Evaluating Robot World Models via Structured Vision-Language Analysis
RoboGaze: Evaluating Robot World Models via Structured Vision-Language Analysis
RoboGaze is a training-free multi-agent VLM evaluator that provides temporally localized diagnostic reports for robot world models. It outperforms existing baselines, including an agentic detector, and improves accuracy in detecting failures in synthetic manipulation videos, achieving 85% mean clean-clip accuracy on the strongest backbone.
RoboGaze introduces a novel, training-free multi-agent VLM evaluator for assessing robot world models, offering temporally localized diagnostic reports with 6 dimensions and 30 failure types. It outperforms existing baselines, including an agentic detector, and significantly improves accuracy in detecting failures in synthetic manipulation videos.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
Source：arXiv Robotics — research abstracts · arxiv.org