Research On Human-Robot Cooperative Grasping Model And Algorithm And Its Application In Building Block Game

In recent years,there has been a rapid development in artificial intelligence technology,sensor technology,and robot control technology,the demand for the intelligence degree of robots continues to increase.Whether in industrial scenarios which pursuing efficiency or household scenarios which pursuing speed,people always hope to achieve closer cooperation with robots.Human-robot collaborative grasping mainly involves two types of robot behaviors:robot delivering objects to humans and robot grasping objects from humans.At present,many robots have achieved the first behavior through various methods,while research on the second behavior is still rare.Robot grasping objects from human hands is a complex human-robot cooperation problem.Robots not only need to consider objective factors such as the location and pose of objects but also need to pay attention to subjective factors such as human intentions and decisions during grasping process.These subjective factors are uncontrollable and difficult to perceive.These subjective factors are more difficult to control and perceive.Therefore,this thesis focuses on the key technical problems of human-robot collaborative grasping,especially the challenging problem of robots using mechanical claws to safely grasp objects from human hands,constructing algorithms and models for human-robot cooperative grasping and applies it to a game system of human-robot building blocks.The main research content of this thesis is summarized as follows:(1)A deep reinforcement learning robot motion strategy RGRL that combines scene simulation and domain randomization is proposed in this thesis.Due to the risk of collision with human hands during the process of the robot moving from its initial position to the grasping position,it is necessary to ensure that the robot’s movement path will not cause harm to human hands.This strategy first simulates the real scene on the computer,then uses domain randomization to automatically generate samples in the simulation scene for the robot to learn,and finally applies the learned parameters to the real scene to make the robot capable of safely moving from the initial position to the grasping position.(2)A human-robot collaborative grasping algorithm G&AR based on gaze and augmented reality is proposed in this thesis.This algorithm addresses the problem of people being unable to understand which position the robot will grasp when there are multiple grasping positions for objects in their hands,which leading to a lack of effective collaboration and communication between humans and robots,indirectly reducing people’s sense of security and trust in robots.G&AR firstly detects multiple candidate grasp positions through RGB images,then presents these positions in front of humans through augmented reality,and finally,people can choose the best grasp position they think from all candidate grasp positions through gaze to achieve collaborative decision-making between humans and robots on the grasp position.(3)An intent understanding algorithm based on multimodal reinforcement learning is proposed in this thesis,and implement a prototype system MRLC for human-robot collaborative building block games.To address the differences in the ways in which different individuals express the same intention,the algorithm integrates reinforcement learning into the process of human-robot collaboration,continuously learning people’s expression habits during the collaboration process,achieving elastic intention understanding.Finally,this thesis applies the above research contents to the prototype system of human robot collaborative building block games,and tests the functionality and performance of the system.