Learning to Manipulate with Weak Supervision

This thesis studies how robots can learn manipulation skills when data and system knowledge are limited. Many robotic methods require accurate models of the robot and large training datasets, which are often unavailable in practice. This work shows that robots can still learn useful behaviours using weaker forms of supervision, such as a single demonstration or unlabelled data. By introducing task-aligned learning structures, robots can apply the same skills across different robot designs and many objects. The results demonstrate that robots can learn to reach, grasp, and manipulate objects with far less data and prior knowledge than traditional approaches require.