Learning quasi-periodic robot motions from demonstration

The goal of Learning from Demonstration is to automatically transfer the skill knowledge from human to robot. Current researches focus on the problem of modeling aperiodic/periodic robot motions and extracting dynamic task parameters from the recorded sensory information. However, it is still not adequate for describing complex behaviors in an unstructured environment, such as searching for an unknown fitting position or painting/polishing an irregular surface. The quasi-periodic and stochastic properties cause a high demand for generalization ability of the modeling techniques. This paper proposes a systematic framework for learning quasi-periodic robot motions, which contains three steps: decomposition, modeling, and synthesization. Firstly FFT transform is performed to identify all the frequencies in the quasi-periodic motion. Then the motion is decomposed into an offset component, a series of harmonic and corresponding envelop components based on the concept of equivalent transformation. The offset component is extracted by Empirical Mode Decomposition, harmonic is separated by notch filter, and envelope component is extracted by Hilbert Transform. These components are either periodic or aperiodic. The aperiodic motions can be modeled by conventional techniques such as Gaussian Mixture Model and recovered by Gaussian Mixture Regression. The periodic motions are modeled in closed-form expressions. Finally, they are synthesized together to regenerate the robot motion. This modeling process captures both the aperiodicity and periodicity of a quasi-periodic motion. Simulation and experiment show that the proposed methods are feasible, effective and can predict robot motions beyond demonstrations. With this generalization ability, it is able to reduce the programming difficulty and demonstration complexity.