HomeBody: A Humanoid That Explores, Remembers, and Acts on Its Own (tml.stanford.edu)

HomeBody is a new humanoid robot designed to independently explore environments, remember tasks, and act on its own, enhancing the synergy between task reasoning and physical execution. Utilizing a Vision-Language Model (VLM), HomeBody selects specific skills and targets based on its perception of the environment, which includes live ego views, map context, and past observations. This intelligent selection process allows the robot to bypass the need for detailed low-level programming by streamlining skill execution through structured tool calls. The significance of HomeBody lies in its advanced integration of perception and action in robotics, making it an important development for the AI/ML community. It employs sophisticated techniques for task execution, such as depth estimation through stereo images and analytical grasp prediction, which collectively enhance its ability to navigate and interact with real-world objects effectively. Its skills encompass a range of tasks, including precise object placement, navigation with spatial awareness, and complex interactions like drawer opening, marking a substantial leap in autonomous robotics that could influence applications in home assistance, logistics, and beyond.