Researchers led by Manav Kulshrestha and Shreyans Jain have built LAGR, a method that chooses how a robot should grasp an object for a given task, according to a paper on arXiv. A hammer, for example, should be held by the handle for hammering but by the head when handing it to someone.

The team built a dataset of 2.68 million scored grasps across 100,911 pairs of images and tasks, covering 17 kinds of objects. LAGR first proposes grasps that are physically feasible whatever the task, each with a feasibility score, and then a re-ranker guided by language weighs all the proposals together and scores how well each suits the task.

In the authors' tests LAGR beat the next best method by 9.3 percentage points, and a Franka Research 3 robot carried out task-oriented grasps on real objects with it.

Source: arXiv Robotics (cs.RO)

Background

Manav Kulshrestha

Robotics researcher

Robotics researcher.