Sequential Monte Carlo controller that integrates physical consistency and motion knowledge

Abstract

Stochastic modeling of motion primitives is a well-developed approach to representing demonstrated motions. Such models have been used in kinematic motion recognizers and synthesizers due to their compact representation of types of high-dimensional motions. They allow robots to synthesize their own motions that is kinematically similar to the demonstrated motions. However, when the robots execute a task during physically interacting with the environment, the robots are required to control the dynamical properties such as contact forces in the similar fashion to demonstrated executions. To achieve this goal, the stochastic model needs to be extended to represent both the kinematics and the dynamics of the demonstrations, and subsequently synthesize motions with physical consistency according to this representation. In this paper, we propose a novel approach to encoding sequences of joint angles and joint torques into a hidden Markov model (HMM). Subsequently, joint torques that satisfy the physical constraints of the equations of motion are generated from the HMM by using a sampling method. These joint torques enable a robot to perform motions kinematically similar to training demonstrations and to control its contact with the environment. Experiments on a robot arm with seven degrees of freedom demonstrate the validity of the proposed approach.