Dynamic Modelling And Pose Manifold Encoding For Robot Learning From Human-Demonstration

With the development of robotics,applications of industrial robots are upgraded from simple tasks in structured environments to complex tasks in unstructured environments.For new applications,traditional robot programming methods have disadvantages of complicated configuration and long preparation cycle.Learning from demonstration integrates human teaching and robot learning to simplify task programming and improve programming efficiency,which is an effective way for efficient robot programming in unstructured environments.However,learning from demonstration based on physical human-robot interaction still has the following problems.With respect to the human-robot interaction,the ability of perceiving the contact force is important for robots.An alternative approach is to estimate the contact force by using the dynamic model.However,the dynamic regressor is singular,which affects the identification accuracy of dynamic parameters.In addition,the estimation result of the contact force is influenced by the identification precision of the practical dynamic model.In terms of robot learning of continuous and curved trajectories,which contain the information of time,position and orientation,both position and orientation need to be modelled.However,elements of the orientation belong to a non-Euclidean space.Applying the Euclidean metric and statistical learning methods to pose trajectory learning causes several errors in statistics calculation and probabilistic modelling,which finally affects the learning effect of curved trajectories.In view of the above shortcomings,this research is carried out with: dynamic modelling,model identification and human-robot interaction,trajectory feature extraction and pose trajectory learning.First,a Lie group-based nonsingular dynamic modelling approach is proposed to establish an identifiable model.Next,a local identification strategy for the motor coupling dynamic model is proposed to identify the practical model,which is applied to model-based human-robot interaction for human-demonstration.Then,an intrinsic clustering method on the orientation manifold is proposed to extract statistical features of demonstrated trajectories.Finally,an approach of encoding and probabilistic modelling on the pose manifold is proposed for robots to learn continuous and curved trajectories in task space.The main contents of this dissertation are as follows.(1)Lie group-based nonsingular dynamic modelling of robots.Regarding of the singular dynamic regressor,the Lie group-based motion decomposition and generalized inertia mapping approaches are proposed to reconstruct the dynamic model and obtain its residual form.By analyzing relationships between coefficients and relationships between generalized inertias in the residual model,criteria for discriminating and regrouping redundant parameters are proposed.This simplifies and unifies the analysis of redundant parameters for different structures and kinematic pairs.Experimental results of nonsingular dynamic modelling and calculation show that,compared with the traditional analytical method for determining identifiable parameters,the computational efficiency of the proposed discriminant and regrouping method is improved by 2 orders of magnitude.Through driving torque estimation experiments,the proposed nonsingular dynamic modelling method is also verified.(2)Identification of motor coupling dynamic model for physical human-robot interaction.To deal with ill-conditioned regressor caused by the insufficient excitation,a local identification strategy is proposed for the modified independent dynamic model.In addition,a practical motor coupling dynamic model is proposed by analyzing the relationship between the transmission system and torques.Meanwhile,the identified model is used for estimating the contact force,which is applied to the impedance model for physical human-robot interaction.Experimental results of dynamic identification and human-robot interaction show that,compared with the traditional identification method,the torque estimation accuracy of the proposed local identification strategy is improved by 36%.Compared with the dynamic model neglecting motor couplings,the proposed motor coupling dynamic model improves the torque estimation accuracy by 32%.(3)Feature extraction on the orientation manifold for trajectories demonstrated through human-robot interaction.An intrinsic representation of geodesic distance on orientation manifolds is proposed by analyzing the geometric characteristics of homogeneous manifolds.Then,on basis of the intrinsic distance and the abstract definition of orthogonal projection in the Euclidean space,a generic geodesic orthogonality and projection method is proposed to reduce the dimension of manifold datasets.Finally,an intrinsic clustering method is proposed on homogeneous manifolds and statistical methods are also proposed for learning clustering number and initializing centers.Experimental results of pose feature extraction show that,compared with the linearization projection methods,the proposed intrinsic clustering algorithm improves the clustering accuracy by 27%.The convergence speed of the proposed statistical initialization algorithm is increased by 36% compared with the random initialization algorithm.As for learning the clustering number,the proposed geodesic projection method improves the accuracy by43%.(4)Pose encoding for robot learning of curved trajectories.Aiming at probabilistic modelling of pose trajectories,this research first analyzes the tangent space transformation and the statistics calculation of Riemannian Gaussian distribution based on the manifold optimization and parallel transport.Then,a novel Riemannian Gaussian mixture model is presented for pose encoding and trajectory learning,which is extended to the product manifold with the product metric.Finally,a manifold Bayesian learning algorithm is proposed to deal with datasets on product manifolds by considering weighted metrics.Based on the tangent space transformation,a manifold Bayesian dynamical model is also presented.Experimental results of robot learning of continuous and curved trajectories show that,with regard to the orientation information,the proposed manifold Bayesian learning framework improves the regression accuracy by 54%compared with the traditional Bayesian learning framework.After modifying the encoding method of the pose manifold,the regression accuracy of the proposed Riemannian Gaussian mixture model is increased by 12%.Both of the proposed probability algorithms can be successfully applied to the robot learning of continuous and curved trajectories.