Research On The Key Technologies Of Service Robot Object Manipulation For Human-robot Handover

As human assistants,service robots are expected to be able to deliver objects with humans mutually in a real environment.At present,in the human-robot handover scenario,service robots are facing problems such as poor generalization of grasping pose detection for unknown objects,low ability to adjust the grasping trajectory of dynamic targets in the scene,and lack of affordance constraints on the grasping action.It is difficult to meet the safety and comfort requirements in the process of human-robot handover.In view of the above problems,this paper conducts research on the key technologies of objects manipulation in the human-robot handover scenario from three aspects:enhancing the robot’s grasping pose evaluation ability,improving the robot’s grasping trajectory response ability,and optimizing the handover planning method.Then we propose an object grasping pose detection method based on 7-DoF grasping representation for human-robot handover,a target-driven reactive closed-loop object grasping method of robot,and a task-oriented heuristic object handover method for robot.The main research content of this paper is as follows:(1)Aiming at the problems of unreasonable generation of grasping candidates and poor generalization of unknown objects in the human-robot handover scenario,a 7-DoF object grasping pose detection method for the human-robot handover scenario is proposed.This method first uses a region-aware grasp detection network to generate reliable 7-DoF grasp candidates for target in the scene,and the mixed sampling and multi-scale feature fusion are used to further improve the detection performance of the grasp detection network and generalization ability.Secondly,by analyzing the force model of the local area around the target’s grasping point,an improved Grasp Affinity Fields(GAFs)calculation method is studied,and the corresponding voxel-level GAFs and view-level GAFs labeling to improve the grasp detection network’s ability to evaluate grasp key points and grasp perspectives.Finally,on this basis,the impact of human hand sampling points in the target area on the graspability area is evaluated by using space division,which effectively improves the security of the grasp candidate for the human hand area.Experimental results show that this method can reasonably generate reliable grasping poses for unknown objects,and effectively avoid the area where human hands are located.(2)Aiming at the poor responsiveness of the robot to the change of the target object’s pose and the low safety of the handover process in H2R(Human-to-Robot)handover scenario,a target-driven robot reactive closed-loop object grasping method is proposed.The method is based on a goal-conditioned and task-auxiliary deep deterministic policy gradient(DDPG)reinforcement learning model to adjust the robot’s end grasping trajectory.The dataset aggregation and behavior cloning are performed on the agent model during the interaction process through a covariant hamiltonian optimization method for motion planning to obtain the collision-free expert experience in control strategy which train and guide the agent.Finally,on this basis,a reinforcement learning model migration method for real robots is proposed.By constructing an environmental octree map and grasping hindsight goals,the reinforcement learning model in real scenes is finetuned,which improves the human-robot performance of the grasping ability and response ability of the robot in the scene.The experimental results show that this method can effectively improve the robot’s ability to grasp the dynamic delivery target,and enable the robot to complete the grasping task safely and efficiently.(3)Aiming at the problem that the delivery planning of robots in the R2H(Robot-toHuman)handover scenario is difficult to meet the comfort requirements,a task-oriented heuristic object handover method for robot is proposed.First,a grasp affordance evaluation model under task constraints is constructed based on a graph neural network.The model learns initial grasp poses for objects under different tasks by inductive knowledge graphs.On this basis,a heuristic-guided robot handover planning strategy and execution strategy is proposed,and the minimum cost model layered optimization is used as a robot heuristic-guided process.The transfer process is optimized from the perspectives of grasping pose affordance and transfer pose rationality.The experimental results show that this method can effectively select the appropriate initial grasping pose for different tasks,and adjust the object handover pose according to the posture of the human hand during the delivery process,thereby improving the comfort of humans during the object handover process.