DescriptionResearchers have been seeking intelligent robotic systems that can accomplish complex tasks autonomously with very little human effort. With the recent progress from both planning algorithms and learning based methods, many low-level primitives can be fulfilled with remarkable quality. This thesis aims at the next step, compositional manipulation tasks. A compositional manipulation task consists of multiple sub-tasks and each sub-tasks can be completed by low-level manipulation, it requires a high-level reasoning to schedule sub-tasks.
In order to reduce labeling effort, we concentrate on self-supervised learning methods. To be specific, we study reinforcement learning algorithms and imitation learning algorithms. Reinforcement learning algorithms explore in the environment and findoptimal behavior based on rewards. As the only feedback is a reward and there is no direct information about sub-tasks provided, we utilize first causality in training predictive model. Another proposed method is to construct a finite-state machine for describing transitions between sub-tasks and include it in the policy input.
Another line of works covered by this thesis are imitation learning methods where the robot is given a collection of human demonstrations rather than the reward. These demonstrations always present a complete task rather than sub-tasks, labels on sub-tasks switching are never provided. We propose a decomposition of the policy, and maximization of demonstration trajectory likelihood based on this decomposition learns the sub-tasks switching autonomously. Finally, we investigate generalizing this manipulation algorithm to unseen objects by removing the requirement of semantic labels on objects. The proposed method describes objects by feature vectors including appearance and shape information, so similar objects share similar features.