Learn to Grasp Unknown-Adjacent Objects for Sequential Robotic Manipulation

Abdulrahman Al-Shanoon

doi:10.1007/s10846-022-01702-4

2022-08-01

PRE

Abdulrahman Al-Shanoon

Haoxiang Lang *

DOI：10.1007/s10846-022-01702-4

原文链接

原文求助

摘要

En 中文

Grasping unfamiliar-adjacent objects based on limited previous information is a daunting task in robotic manipulation. It is substantially more difficult to grasp an object in such a scenario than grasping secluded objects. For this reason, recent solutions typically require non-prehensile actions prior to grasping (e.g., pushing, toppling, squeezing or rolling). However, these solutions play a loose role and causing delays. The non-prehensile action should have intended utility and effect on the consecutive grasping action, because it is a sequential decision-making problem. This paper takes a step towards solving the issue by introducing a self-learning strategy to manipulate unknown objects in challenging scenarios based on minimal prior knowledge. The developed system learns jointly pre-grasping (non-prehensile shifting) and grasping (prehensile) actions using model-free deep reinforcement learning. The agent comprehends sequences of pregrasp manipulations that purposely lead to successful potential grasps. The system is object-agnostic, which does not require task-specific training data or predefined object information (e.g., pose estimation or 3D CAD models). The proposed model trains end-to-end policies (from only visual observations to decisions-making) to seek optimal manipulating strategies. Perception network maps visual inputs to actions, as dense pixel-wise Q-values, and learns quickly through trial-and-error manner. Experimentation findings have demonstrated the effectiveness of the joint learning between pregrasp manipulation and grasp policies, in which the success rate of grasping has greatly increased. The proposed system has been experimentally tested and validated in simulations and real-world settings using 6DOF robot with two-finger gripper.

Keyword:

Self-supervised joint-learning

Deep reinforcement learning

Pre-grasping actions for sequential robotic manipulation