Enabling robots to perform complex manipulation tasks in open-world environments requires approaches to intelligence that bridge the traditional divide between symbolic and subsymbolic systems. Hybrid systems that combine reasoning over symbolic knowledge with probabilistically approximately correct inference over probability distributions promise to make robot manipulation actions generalizable, interpretable, and scalable to address real-world challenges. This

4D computer vision using egocentric visual sensors on head-mounted devices (such as Aria glasses) is gaining momentum. While substantial progress has been made in recent years, various open challenges remain in the field, including robust 3D human and hand pose estimation from egocentric cameras, optimal egocentric camera placement, real-time processing, and fulfilling ergonomic and power