Online Multi-modal Imitation Learning via Lifelong Intention Encoding

For the imitation learning of robotic manipulation, generalized ability that enable robots to continuously acquire new tasks and reuse pre-existing experience from previous learned tasks is necessary. In this paper, we aim to learn multi-modal imitation demonstrations with different intentions by optimizing a sparse coding lifelong intention dictionary, which enable robots to autonomously imitate various complex behaviours. We investigate multi-intention which encode from an action imitation learning tasks sequence that comes from various experts with diverse intention. Our approach developed a lifelong imitation learning framework for inverse reinforcement learning (IRL) to deal with multi-intention imitation learning problem. suitable for inverse reinforcement learning intention encoding from open-ended demonstrations. Our work is aimed at producing the reward function of general tasks and maintaining a reusable experience dictionary automatically which can be used for online multi-intention sparse encoding.