Researchers Of The University of Utah Develops A New Supervised Learning Approach To Grasp Planning In Robots

Artificial Intelligence News

Researchers-Of-The_University-of-Utah-Develops-A-New-Supervised-Learning-Approach-To-Grasp-Planning-In-Robots Researchers Of The University of Utah Develops A New Supervised Learning Approach To Grasp Planning In RobotsA team of researchers at the University of Utah have recently formed a probabilistic grasp planner that can explicitly create grasp types to plan high-quality accuracy and power grasps in real time. Outlined in a paper, their supervised learning approach can effectively plan both power and accuracy grasps for a dealt object.

Grasping different types of manipulation tasks, it requires different sorts of grasps, for both humans and robots. For instance, carrying a heavy tool like a hammer requires a multi-fingered power grasp which gives stability, while carrying a pen requires a multi-fingered precision grasp, as this can impart dexterity on the object. The University of Utah research team, when testing their earlier approach for grasp planning, noticed that it almost always produced power grasps in which the robot’s hand wraps around an object, with large contact regions between its hand and the object. These grasps are constructive for completing a range of robotic tasks like picking up objects somewhere else in the room, but yet they are obstructive when performing in-hand manipulation tasks. Devised by Tucker Hermans, one of the researchers who carried out the study and his colleague Qingkai Lu, the approach to grasp planning in which a robot learns to prognosticate grasp success from past experiences. The robot tries various sorts of grasps on different objects, for recording which of these were successful and which failed. This data is then utilized to train a classifier to foresee whether a given grasp will succeed or not.

The supervised learning approach can plan different kinds of grasps for previously unseen objects, even when only partial visual information is available. The researchers assessed their model and compared it with a model that does not encode grasp type. Their findings then recommend that modeling grasp type can better the success rate of generated grasps, with their model exceeding the other method. The Hermans and Lu’s approach could support the development of robots that can create a diverse set of grasps. Ultimately, it would enable these robots to complete a larger variety of tasks that involve different types of object manipulation.