Motion synthesis from stochastically encoded motion primitives for anthropomorphic robotic arm

摘要

Imitation learning is an efficient framework that allows anthropomorphic robotic arms to synthesize human like movements. However, the synthesis of the exactly same movements as learned ones is not helpful in real environments since the environments are different from the one where the movements were learned. For instance, a robotic arm learns the movement of manipulating an object by observing a human manipulating the object. This movement that is precisely same as learned one cannot be reused in the new situations since the the locations of the object are varied. Therefore, synthesis of a movement that not only looks natural and but also is adaptive to the new environment is necessary for the anthropomorphic robotic arm that completes a specific task. This paper describes a novel approach to synthesizing natural movements for the anthropomorphic robotic arm based on the maintained demonstrations given by its human partners. The anthropomorphic robotic arm is trained through kinesthetic demonstrations by its human partner, encodes the demonstrations into stochastic models, and synthesizes new movements dependent on the current environment. The proposed approach designs the objective function which evaluates the similarity between the synthesized movement and maintained demonstration, and the satisfaction of being kinematically constrained to the environment. The movement that maximizes this objective function can be found, and consequently the anthropomorphic robotic arm can reproduce the adaptive movement to the current environment such that it can accomplish a specific task.