Self-supervising 3D Scene Representations for Perception and Control

Explore cutting-edge research on self-supervised 3D scene representations in this hour-long robotics colloquium lecture delivered by Katerina Fragkiadaki from Carnegie Mellon University. Delve into the limitations of current CNN models and discover novel approaches that enable machines to understand fundamental concepts about objects and their 3D properties. Learn how these advanced models can map 2D and 2.5D inputs into comprehensive 3D feature maps, demonstrating object permanence, emergence without human annotations, and generalized object dynamics. Gain insights into the potential applications of these techniques in language grounding and visual simulations. Understand the speaker's background in machine learning, computer vision, and AI, including her work on developing algorithms for mobile computer vision and physics-based reasoning for interactive agents.