Robot Multimodal Introspection And Learning Based On Nonparametric Bayesian Model

Human-robot coexistence and cooperation have become an urgent demand in modern manufacturing and services industries,which result in the concept of "Human-Robot Collaboration,HRC" and "Human-Robot Interaction,HRI" have naturally emerged for many years.Sharing intelligence,having a common behavior,and cooperating with people to accomplish common tasks(behavior,task,and intelligence)are their basic characteristics and elements,also have become the consensus of domestic and foreign academic circles.The prerequisite of HRC is human-robot peaceful coexistence,that is,"safety collaboration".Although many stable and robust robot control algorithms exist,internal modeling errors or external disturbances still affect the robotic systems,such as human collision,object sliding,and tool collision.The HRI scenarios are the unstructured and non-standardized dynamic environment that cannot be completely modeled and analyzed.To endow robots with longer-term autonomy and safer HRC environment,it is necessary to model the real-time multi-modal signals for implementing the accurate behavioral perception(Introspection)and anomaly recovery policy learning.In this dissertation,the robot multimodal introspection and learning are theoretically investigated and the main contents and achievements are as follows:(1)To address the problem of learning and generalizing complex tasks of robots,a method of combining Dynamic Movement Primitive(DMP)and Finite State Machine(FSM)is proposed in this dissertation,which divides the robot complex manipulation task into sequential motion primitives for improving the adaptability and diversity of the manipulation tasks,i.e.parameterized directed graph descriptions.The proposed method is based on the basic theory of robot learning from demonstration.(2)To address the problem of multi-modal fusion,this dissertation abstracts its research object as how to effectively establish and interpret the probabilistic model of multivariate time series.The uncertainty of the number of hidden states and the high transition frequency between hidden states are two critical problems of Hidden Markov Model(HMM)in modeling multivariate time series,which will greatly weaken the modeling performance and time consistency.Based on this,several nonparametric Bayesian models named Sticky Hierarchical Dirichlet Process Hidden Markov Models(sHDP-HMM)are investigated to jointly model the robot end-effector's velocity,force/torque,tactile signals of as well as their related statistical information(such as mean and variance)during robot manipulation task.(3)With the multi-modal fusion,nonparametric Bayesian models-based methods for robot real-time motion behavior recognition and anomaly monitoring are proposed in this dissertation.First,the normal multi-modal signals of each movement primitive are modeled using sHDP-HMM with the help of a parametric description of the robot's manipulation task.Then,the robot behavioral recognition is implemented by comparing the cumulative log-likelihood values of real-time observations.Finally,three types of abnormal thresholds for robot abnormal monitoring are proposed when the behavior is known,which including log-likelihood value,log-likelihood gradient value as well as mapping relationship between latent state and log-likelihood value.(4)With the multi-modal anomaly monitoring,nonparametric Bayesian models-based multi-objective classifier for multi-modal anomaly classification are proposed.Detailly,anomalous samples are extracted before and after the occurrence of anomaly events according to the size of the given window,sHDP-HMM model is learned for each anomaly type individually,and the optimal model is selected by cross-validation method;anomaly classification is tackled by comparing the cumulative log-likelihood values of testing samples cross all learned models.(5)With the anomaly monitoring and classification,two task-level robot anomaly recovery policies are respectively proposed by learning the experience and intention of human treating accidental and persistent anomaly events,which including the re-enactment policy for accidental anomalies by using a multinomial distribution as well as the adaptation policy for persistent anomalies by the parametric representation of human's demonstrations.(6)Taking all the research contents and achievements of this dissertation into consideration,a multi-modal fusion based robot introspective and learning system framework named SPAIR(Sense-Plan-Act-Introspect-Recover)is proposed based on the traditional robot control framework Sense-Plan-Act(SPA),which adds the phases of Robot Introspection(behavior recognition,anomaly monitoring,anomaly classification)and anomaly recovery.The framework intentionally endows the robot with longer-term autonomy and safer human-robot collabrative environment,which mainly includes four functional modules:1)directed graph representation of complex robot manipulation tasks;2)robot movement generation and generalization learning;3)real-time anomaly monitoring and classification during robot execution;4)robot anomaly recovery policy learning.