Doctoral Thesis: Optimistic Active Learning of Action Models for Robotic Manipulation

Caris Moses

Abstract:

Manipulation tasks such as construction and assembly require reasoning over complex object interactions. In order for a robot to successfully plan for, execute, and achieve a given task these interactions must be modeled accurately. Existing methods for engineering these dynamics models fail to accurately capture underlying complexities such as friction or non-uniform mass distribution. Therefore, in this work we leverage a data-driven approach to acquiring action models. We propose active learning strategies which aid the robot in learning action models efficiently, with the ultimate goal of using them with a planner. Additionally, we supply the robot with initial optimistic action models which are a relaxation of the true unknown transition model and are easier to specify that fully accurate action models. We are generally interested in the scenario in which a robot is given an initial (optimistic) action model, an active learning strategy, and a space of domain-specific problems to generalize over. In this talk I will present several active learning strategies for manipulation tasks which leverage optimism. I will also give results in several domains involving complex object interactions such as constrained mechanisms, block stacking, and tool use.

Details

• Date: Friday, April 22
• Time: 3:00 pm
• Location: 32-G449 (Kiva)

Additional Location Details:

Thesis Supervisors: Professors Leslie Pack Kaelbling and Tomás Lozano-Pérez