A new study from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) demonstrates a significant advancement in robotic dexterity. Researchers have developed a system that enables a robot hand to manipulate over 2,000 diverse objects, from mugs and plates to items as delicate as a single grape, using just one human demonstration. The system, named ‘Dexterity from Demonstration’ (DexDem), relies on a camera-equipped robotic hand and a neural network that learns generalizable skills from the initial demonstration. This approach allows the robot to adapt its grip and movements to handle objects of different shapes, sizes, and weights, and to recover from disturbances like accidentally dropping an item. The technology represents a step toward more adaptable robots capable of performing complex tasks in unstructured environments like homes and warehouses. Read the full article at: https://technologyreview.com/2024/05/20/1093035/robot-hand-learns-object-manipulation-from-one-demonstration/