Google’s DeepMind team recently introduced Open X-Embodiment, a collaborative effort with 33 research institutes, akin to the influential ImageNet database. This new resource aims to propel robotics research by providing a diverse dataset of over 500 skills and 150,000 tasks from 22 different robot embodiments.
Vincent Vanhoucke, Google DeepMind’s Head of Robotics, highlights the pivotal role of perception technology in enabling real-world robotics. His team’s focus on general-purpose methods rather than specific embodiments sets the stage for a new era in robotics. Vanhoucke emphasizes the close collaboration between Everyday Robots and DeepMind, where they pioneered machine learning-based robotic grasping.
While the transition to general-purpose robotics is ongoing, Vanhoucke underscores the importance of technology evolution. He mentions the recent success of the RT-1-X project, demonstrating the feasibility of general approaches independent of specific embodiments.
Generative AI emerges as a central player in robotics, providing common sense reasoning and an understanding of everyday concepts. Large language models offer valuable insights, enabling applications in robot planning, interactions, and manipulations. Simulation, while a key component, also presents challenges in bridging the gap to reality. Advances in generative AI address this issue, potentially revolutionizing how robots perceive and interact with the world.