Trustworthy Visual Predicates for Robust Manipulation Understanding under Degradation
Researchers have developed a new framework to assess the reliability of visual predicates used in understanding robotic manipulation. This framework evaluates how well predicates like contact, support, and grasp perform under various degradation conditions such as blur, occlusion, and frame dropping. Experiments on several datasets demonstrated that while static predicates are relatively robust, dynamic and derived predicates are more susceptible to errors, significantly impacting downstream manipulation understanding accuracy. AI
IMPACT Provides a diagnostic layer for improving robotic manipulation understanding by identifying weaknesses in visual predicate recognition under degraded conditions.