A Machine Learning Algorithm for Collaborative Push-and-Shove Boosts Robots Grasping Accuracy
By breaking a pile of objects apart with a few careful shoves, a robot arm's ability to grasp objects goes from 35 percent to 97 percent.
A team of researchers from Hangzhou Dianzi University and the University of Sydney has developed a machine learning system, which gives robot arms the ability to push and pull objects autonomously — breaking stacks to separate them before grasping individual objects.
"Directly grasping the tightly stacked objects may cause collisions and result in failures, degenerating the functionality of robotic arms," the team explains in the abstract to the study. "Inspired by the observation that first pushing objects to a state of mutual separation and then grasping them individually can effectively increase the success rate, we devise a novel deep Q-learning framework to achieve collaborative pushing and grasping."
"By creating a unique collaborative pushing and grasping reinforcement learning network and reward functions, the robot can effectively grab and sort closely packed objects in real-world circumstances," explains Jing Zhang, research fellow at the University of Sydney and co-author of the study. "Industrial parts sorting and residential waste sorting are two examples of how the proposed technology could be used."
The trick lies in not simply having the robot arm attempt to grasp at an individual object straight off the bat, which showed a mere 35 percent success rate in testing; instead, it pushes at a pile of objects to separate them before attempting to grasp any individual object, an approach that beat 97 percent accuracy on average during real-world testing — hitting as high as 100 percent in simulations, where pushed objects couldn't move out of the robot arm's reach.
"Experimental results demonstrate the superiority of the proposed method over the non-RL [Reinforcement Learning] pushing method and directly grasping method for this challenging task," the researchers concluded, "as well as its fast learning speed, good generalization capability and robustness."
The team's work has been published in the IEEE/CAA Journal of Automatica Sinica under open-access terms.