Grasp-TAGS picks the grasp that best suits a mobile robot's rearrangement task

Xun Tu, Minghao Zou and Karthik Desingh have described Grasp-TAGS in an arXiv paper, a method for mobile manipulators moving large objects to choose which grasp to use before touching the object. Different stable grasps can lead to very different results later in the task, but learning methods usually assume a fixed starting grasp.

The team trained one reinforcement learning policy across several starting grasps and distilled its critic into a small network that ranks candidate grasps from the grasp and goal alone. Grasp-TAGS reached an average success rate of 89.84%, against 79.87% for the best single-grasp policies, and reduced final orientation errors for all objects tested.