Abstract:
A general purpose robot will need to interact with objects in cluttered environments with minimal supervision. Machine learning provides methods that can deal with these complex tasks without explicitly modelling the environment. More recently, deep learning techniques combined with large scale data has revolutionized the fields of computer vision, language processing and reinforcement learning. The availability of large datasets in these fields has made end-to-end learning approaches feasible. The question at the center of this thesis is how can we obtain and learn from large scale data for robotics.

Unlike the aforementioned fields, robotics involves real hardware systems, which presents three unique challenges for large scale learning. The first challenge we consider is the physical nature of robotics where every datapoint needs to be executed on a real system. We address this with a self-supervised technique where the robot both collects and labels real-world data. This has enabled our robots to grasp and push novel objects in clutter. The next challenge is that robots are slow, which limits the amount of data we can collect. For this, we accelerate robot learning by using adversarial agents that extract useful information and multi-task learning that shares feature representation.

The final challenge we address is that commercially available industrial robots are expensive, which makes parallelizing learning expensive. To tackle this, we present a low-cost hardware platform that allows us to efficiently scale up robot learning to multiple robots working in parallel. Furthermore, we demonstrate that data obtained from homes allows us to learn robust grasping models that work in new homes with novel objects.

More Information

Thesis Committee Members:
Abhinav Gupta, Chair
Christopher Atkeson
Martial Hebert
Siddhartha Srinivasa, University of Washington
Pieter Abbeel, University of California, Berkeley