| Recently,the application scope of robotic arms has expanded to intelligent logistics,laboratories,offices and other scenarios.The diverse application scenarios have raised higher requirements for the autonomous operation planning ability of robots.However,the existing robotic arm operation planning methods have many issues,such as a lack of explicit use of environmental information,low motion planning efficiency,and a lack of outer feedback.The academic community is currently focused on combining perception and robot manipulation planning to solve these problems,with commonly used methods including point cloud autoencoder,learning-based motion planning algorithms,and point cloud-based robotic arm state estimation algorithms,etc.While these methods effectively improved the performance of robotic arm,they still have many unresolved issues.Firstly,point cloud autoencoder used for motion planning have poor generalization.Secondly,the current learning-based motion planning algorithms have low success rates.Thirdly,point cloud autoencoder have inaccurate environmental representation when processing multiobject point clouds.Finally,point cloud-based robotic arm state estimation algorithms are slow in updating,and there is a discrepancy between the state of the end effector and the observation.To solve these problems,this thesis studies the point cloud autoencoder,point cloud-guided motion planning algorithm,and point cloud-based position control algorithm applied in desktop environment,utilizes technologies such as Bayesian learning,multi-mode learning,sequence learning,state estimation,to realize high-efficiency,highsuccess-rate,and high-accuracy robotic arm manipulation planning.This thesis proposes a variational point cloud autoencoder network to address insufficient point cloud autoencoder environment representation extraction and poor generalization ability issues.Firstly,Bayesian learning theory is introduced and a point cloud variational loss function is constructed to construct the factors disentangle and interpolatable hidden space to improve the generalization ability of the point cloud autoencoder.Secondly,a topological point cloud reconstruction network is constructed to enhance the environment representation extraction ability of the point cloud autoencoder.Finally,reconstruction experiments are conducted to verify the generalization and reconstruction ability of the proposed autoencoder.Subsequently,this thesis designs a multi-motion pattern sampling network to improve the planning efficiency and success rate of the point cloud-guided motion planning algorithm.The network is designed as a combination of a coefficient network and a multi-motion pattern network to deal with the discontinuity property,and to learn different motion patterns accurately and autonomously.In addition,a point cloud-guided rapidly-exploring random tree motion planning algorithm is proposed and compared with other motion planning algorithms.For the commonly encountered multi-object environments in motion planning,this thesis proposes a multi-object point cloud Transformer and a separated chamfer distance loss function to solve the problem of inaccurate representation of multi-object point cloud environments.The multi-object point cloud reconstruction problem is modeled as an1-to-N sequence generation problem and proposes a multi-object point cloud Transformer network structure.Subsequently,to address the problem of point cloud misalignment caused by chamfer distance,the separated chamfer distance is proposed,and it is proved that the separated chamfer distance is an upper bound of the chamfer distance.Finally,reconstruction,motion planning,and ablation experiments are carried out to verify the ability of the proposed network in processing multi-object point clouds.This thesis also designs a point cloud-based end-effector state estimation algorithm to achieve efficient and accurate robotic arm outer-loop feedback and designs a position control algorithm based on the estimated end-effector states.First,the point cloud-based end-effector state estimation algorithm is proposed to improve the estimation efficiency and accuracy.Further,the position control and inverse kinematic robotic arm state estimation algorithms are proposed based on the nominal Jacobi matrix to achieve the control and overall state estimation of the robotic arm.Finally,the point cloud-based end-effector state estimation algorithm is compared with other robotic arm state estimation methods and conducts experiments on robot state estimation algorithms and position control.Finally,this thesis builds an efficient and accurate motion system experiment platform for robotic arms based on point clouds applied in desktop environment.Based on the above algorithms,robotic arm manipulation planning experiments are carried out using the step-by-step motion method of the robotic arm.Combining perception and robotic arm motion,this thesis studies the operation manipulation planning of robotic arms in the desktop environment,designs a variational point cloud autoencoder,a multi-object point cloud transformer,and a multi-modal sampling network,proposes a point cloud-guided fast-expansion random tree motion planning algorithm and an efficient position control algorithm based on point clouds,and designs experiments to verify the effectiveness of the above algorithms.Based on the above algorithms,the robotic arm can achieve efficient and high-success-rate manipulation planning and achieve precise simulated manipulations.The neural network structures and algorithms proposed in this thesis provide a theoretical foundation and technical support for autonomous manipulation planning of robotic arms. |
