Offline trained object detectors perform well on public datasets but degrade in the real world due to distributional shifts. Embodied intelligence, by enabling sample collection in the target environment, offers a solution. A key challenge is determining which samples to collect. Current methods primarily focus on semantically uncertain data, overlooking the importance of missed detection samples. This paper addresses this by using a class-agnostic detector to identify potential objects within the agent’s egocentric view, integrating local and global information as rewards. Experiments show that the proposed method effectively captures informative detection samples for perception model training.