It is often useful for robots to actively build a model of an unknown 3D scene to enable tasks such as manipulation, mapping and object search. To do so requires choosing a representation to accumulate spatial knowledge, and strategies for selecting actions to acquire relevant spatial information and interact with objects. To achieve reliable performance, the data representation and planning algorithm should take into account uncertainty in the robot’s belief of the world, to mitigate the effects of sensor noise and promote informative and robust actions.
In this thesis we adopt a spatial representation based on simple geometric solids, such as cylinders and cuboids, and we extend this to probability distributions over the shape parameters. By augmenting the representation with uncertainty, the robot can reason over object-level spatial information about the shape parameters. Our approach enables the shape of novel objects to be inferred online from a sequence of views, and supports predicting viewpoint information and grasp robustness.
Committee: Professor Nicholas Roy, supervisor
Professors Tomas Lozano-Perez, Russ Tedrake
To attend this defense, please contact the doctoral candidate at prentice at mit dot edu