Active Perception and Long-term Understanding of Dynamic Human-Centric Scenes

Explore a technical talk that delves into developing autonomous robots capable of operating in dynamic, human-shared environments. Learn about an innovative autonomy pipeline that addresses the challenges of long-term robot operation, focusing on the detection and representation of complex semantics, short-term motion, and long-term changes in dense robot maps. Discover the Khronos framework and Dynamic Scene Graphs (DSGs), which enable robots to represent semantic symbols and predict future scene evolution based on collected data. Examine how embodied agents can utilize actionable scene representations and predictions to enhance their world models and perception capabilities over time. The presentation demonstrates these methods through real-time implementations on autonomous aerial and ground robots, showcasing practical applications in various fields including home service, industrial inspection, and disaster response. Delivered by MIT SPARK Lab postdoctoral research fellow Lukas Schmid, whose work has earned multiple accolades including the RSS 2024 outstanding systems paper award and the ETH Medal for outstanding PhD Theses.