Robust Embodied IQA: Execution-Oriented Quality Estimation via Semantic Degradation Prompting and Operation-Aware Calibration
Abstract
In Embodied AI, image quality assessment (IQA) should move beyond human visual preference and estimate whether degraded visual inputs remain useful for real-world task execution. Real-world degradations may affect task-relevant visual evidence, multimodal reasoning, and action-related decision making, making conventional perceptual IQA insufficient for task-oriented embodied scenarios. To address these challenges, we propose a degradation-aware framework for task-oriented Embodied IQA. First, to reduce potential feature redundancy and improve multimodal degradation understanding, we introduce Semantic Degradation Prompting. It converts predicted noise types and levels into natural language descriptions, providing explicit degradation context for task-conditioned multimodal quality assessment. Second, to improve quality estimation under limited and long-tailed training data, we design Hierarchical Quality Assessment. It jointly estimates absolute quality scores and relative quality differences, while employing a task-difficulty-aware contrastive loss to enhance representation robustness. Third, to improve prediction reliability under local spatial sensitivity and execution uncertainty, we propose Operation-Aware Calibration. It aggregates absolute-relative ranking disagreements, spatial perturbation responses, including center masking and downsampling, and prediction uncertainty, and then applies monotonic plateau calibration to stabilize unreliable middle-score intervals. Our method ranked first in the Machine-oriented Visual Media Quality Assessment (MoIQA) Track 3, which evaluates predicted quality scores on real-world robot sensor data against binary task execution outcomes, demonstrating its effectiveness for real-world robot-oriented image quality assessment.