Committee Chair: Dmitry Berenson

Abstract:
A central challenge for developing general-purpose robot assistants is the development of algorithms for robot manipulation that can perform a wide range of tasks across a diverse set of environments. This thesis develops planning and trajectory optimization methods that can adapt to new and unforeseen systems. The key to these methods is the ability of robots to learn from experience and reason about related uncertainty. Using modern machine learning and approximate probabilistic inference techniques, the work in this thesis improves the ability of planning methods to do so.

Probabilistic inference is useful in two ways. First, by using a probabilistic framing, probabilities can be used as a way of expressing confidence in our current models. I develop a method that learns to predict the uncertainty of a given dynamics model with a small amount of data collected online and avoids areas where the model is uncertain. I also propose an approach that learns a generative model of control sequences to complete a given task. I demonstrate that we can detect and adapt this generative model to situations where the environment differs from the training environments.

Second, I incorporate probabilistic inference into the proposed methods by viewing planning itself as an inference problem. By framing planning as inference, we construct probability distributions over trajectories. This framework allows me to develop a method that views constrained trajectory optimization as inference, generating diverse sets of constraint-satisfying trajectories for completing manipulation tasks. This allows improved adaptation to online disturbances, since at any given time, there is a set of trajectories to select from. I demonstrate the effectiveness of this method on several different tasks, including a 7DoF manipulator turning a wrench and a 16DoF multi-fingered hand turning a precision screwdriver.