Developing personalized cognitive robots that help with everyday tasks is one of the on-going topics in robotics research. Such robots should have the capability to learn skills and perform tasks in new situations. In this thesis, we study three research problems to explore the learning methods of robots in the setting of manipulation tasks. In the first problem, we investigate hand movement learning from human demonstrations. For practical purposes, we propose a system for learning hand actions from markerless demonstrations, which are captured using the Kinect sensor. The algorithm autonomously segments an example trajectory into multiple action units, each described by a movement primitive, and forms a task-specific model. With that, similar movements for different scenarios can be generated, and performed on Baxter Robots.

The second problem aims to address learning robot movement adaptation under various environmental constraints. A common approach is to adopt motion primitives to generate target motions from demonstrations. However, their generalization capability is weak for novel environments. Additionally, traditional motion generation methods do not consider versatile constraints from different users, tasks, and environments. In this work, we propose a co-active learning framework for learning to adapt the movement of robot end-effectors for manipulation tasks. It is designed to adapt the original imitation trajectories, which are learned from demonstrations, to novel situations with different constraints. The framework also considers user feedback towards the adapted trajectories, and it learns to adapt movement through human-in-the-loop interactions. Experiments on a humanoid platform validate the effectiveness of our approach.

In order to further adapt robots to perform more complex manipulation tasks, as the third problem, we are investigating a framework that the robot could not only plan and execute the sequential task in a new environment, but also refine its actions by learning subgoals through re-planning/re-execution during the practice. A sequential task is naturally considered as a sequence of pre-learned action primitives, each action primitive has its own goal parameters corresponding to the subgoal. We propose a system to learn the subgoals distribution of given task model using reinforcement learning by iteratively updating the parameters in the trials. As a result, by considering the learned subgoals distribution in sequential motion planning, the proposed framework could adaptively select better subgoals to generate movements for robot to execute the task successfully. We implement the framework for the task of ''openning a microwave'' involving a sequence of primitive actions and subgoals and validate it on Baxter platform.