Columbia Engineers have developed an innovative robot hand that combines advanced touch sensing with motor-learning algorithms, enabling manipulation of objects without reliance on vision. This ground-breaking technology allows the robot to perform complex tasks such as rotating unevenly shaped objects in its grasp while maintaining stability—all without visual feedback. By utilizing tactile sensors and machine learning, the robot can operate effectively in various lighting conditions, including total darkness, which is a significant advancement over previous methods that relied on visual data.
The robot hand features five fingers with 15 independently actuated joints, each equipped with cutting-edge touch-sensing technology. Through deep reinforcement learning, the robot underwent extensive simulation training, equating to about a year of practice condensed into hours, before transferring its skills to real-world applications. The researchers envision applications in logistics, manufacturing, and assistive robotics, with the ultimate goal of merging abstract intelligence with embodied dexterity. This dual approach aims to enable robots to execute complex physical tasks, such as making sandwiches, by leveraging both semantic understanding and fine motor skills—marking a significant milestone toward achieving true robotic dexterity comparable to human capabilities.
