Abstract:
Real world robots need to continuously learn new manipulation tasks in a lifelong learning manner. These new tasks often share sub-structures (in the form of sub-tasks, controllers) with previously learned tasks. To utilize these shared sub-structures, we explore a compositional and object-centric approach to learn manipulation tasks. While compositionality in robot manipulation can manifest in different ways (e.g. dynamics, policies), in this thesis we focus on compositionality for learning the task structure. Overall, our approach is more robust to task variations and allows for sample efficient learning while our lifelong learning approach allows new tasks to be learned efficiently.

Among our completed works, we first focus on compositional preconditions. We show how complex manipulation tasks, with multiple objects, can be simplified by focusing on pairwise object relations. We use a self-supervised learning objective to learn an object-relation model, which we then use to efficiently learn manipulation skill preconditions. Our next work composes policies using object-centric task-axes controllers. Our task-axes controllers learn the skill structure and are composed into specialized policy representations for individual tasks. These representations are robust to environment variations and are learned from limited data. We then focus on compositionality for continual task learning. For this we first propose skill effect models, which predict the effects of stereotypical skill executions. We utilize skill modularity to learn skill effects for an increasing number of tasks over time.

For our proposed works, we first investigate the role of biases in robot learning. Specifically, we look at how we can design different inductive biases and integrate them in manipulation learning using appropriate manipulation primitives. Finally, by utilizing the recent advancement in robot learning, we propose to combine different task representations at different hierarchies, allowing the robot to reason at multiple levels and learn manipulation tasks more efficiently.

Thesis Committee Members:
Oliver Kroemer, Chair
Abhinav Gupta
David Held
Dieter Fox, University of Washington

More Information