Programming by demonstration, also called "imitation learning," offers the possibility of flexible, easily modifiable robotic systems. Full-fledged robotic imitation learning comprises many difficult subtasks. However, we argue that, at its core, imitation learning reduces to a regression problem. We propose a two-step framework in which an imitating agent first performs a regression from a high-dimensional observation space to a low-dimensional latent variable space. In the second step, the agent performs a regression from the latent variable space to a high-dimensional space representing degrees of freedom of its motor system. We demonstrate the validity of the approach by learning to map motion capture data from human actors to a humanoid robot. We also contrast use of several low-dimensional latent variable spaces, each covering a subset of agents' degrees of freedom, with use of a single, higher-dimensional latent variable space. Our findings suggest that compositing several regression models together yields qualitatively better imitation results than using a single, more complex regression model
展开▼