Next-Generation Robot Perception: Hierarchical Representations, Certifiable Algorithms, and Self-Supervised Learning

Explore cutting-edge advancements in robot perception through this colloquium talk by Luca Carlone from MIT. Delve into the development of next-generation perception systems that aim to bridge the gap between robot and human perception capabilities. Learn about hierarchical representations for scalable metric-semantic scene understanding, the concept of 3D scene graphs, and their importance in efficient storage and real-time perception algorithms. Discover the progress in designing certifiable algorithms for robust estimation, including the novel concept of "estimation contracts" that provide performance guarantees for robot perception problems. Examine the relationship between certification and self-supervision, and how this insight is applied to 3D object pose estimation, resulting in self-supervised algorithms that rival fully supervised methods without requiring manual 3D annotations. Gain insights from Carlone's extensive research background and achievements in nonlinear estimation, optimization, and probabilistic inference applied to robotics and autonomous systems.