Spotlighting Task-Relevant Features: Object-Centric Representations for Better Generalization in Robotic Manipulation
The episode discusses a paper titled "Spotlighting Task-Relevant Features: Object-Centric Representations for Better Generalization in Robotic Manipulation." Hosts discuss how Slot-based Object-Centric Representations (SOCRs) improve robotic policy reliability by separating task signals from visual noise. They cover the benefits of this structural approach, limitations like slot merging under clutter, and proposed improvements such as adaptive slot capacity tuning.