Synthesis of kinematically constrained full-body motion from stochastic motion model

Humanoid robots are expected not only to understand human behaviors but also to perform human-like actions in order to be integrated into our daily lives. Learning by imitation is a powerful framework that can allow robots to generate the same motions that humans do. However, generation of motions by robots that are precisely the same as learned motions is not typically helpful in real environments, which are likely to be different from the environment where the motions were learned. For example, a robot may learn to reach for a glass on a table, but this motion cannot be used as is to reach for a cup on a cupboard shelf because the location of the cup is different from the location of the glass. Objects manipulated by robots and humans can be located in a variety of places. The robots therefore have to synthesize motions that depend on the current environment in order to reach a target object. Adaptive motion synthesis from memorized motions is an essential technique for allowing robots to perform human-like motions and accomplish motion tasks. This paper proposes a novel approach to synthesize full body motion by using both motions encoded as Hidden Markov Models and kinematic task constraints. We design an objective function that evaluates similarity between synthesized and memorized motions, satisfaction of the kinematic constraints, and smoothness of the generated motion. We develop an algorithm to find a motion trajectory that maximizes this objective function. The experiments demonstrate the utility of the proposed framework for the synthesis of full body motions by robots.