Robotic Materials: Manipulation for the Real World
In order to be successful at a task, a robot must have a rich understanding of its environment and an effective way of interacting it with it. While deep learning for robotic manipulation plays a pivotal role in exploiting information encoded in the environment, its capabilities are ultimately limited by its sensing modalities and its end effector. However, most practitioners do not take these into consideration while tackling a task. For most task vision and motor encoding, which are the most common sensing modalities used in practice, are insufficient. So, in order to take steps towards solving general manipulation, we first took a look at what information is relevant to robotic manipulation and how to exploit it. We then looked at how to make manipulation easier. As a result of our efforts, we came up with a novel smart robotic gripper that provides rich sources of data important to a wide range of manipulation, all while incorporating mechanisms relax constraints imposed by the environment. The first part of this talk takes an in-depth look at our smart gripper and discusses how we believe that it will make deep learning more efficient.
The second part of the talk emphasizes our approach to empowering users, not the users as in the system integrators or the developers but the users as in the people who will be working alongside the robot. The problem with deep learning is that it is essentially a black box, understanding what the program is doing is almost impossible. This highly restricts businesses or consumers using the robot, because it is unintuitive and cumbersome to repurpose the robot for another task or update the current task if trained with deep learning. So we put a priority to interpretability. We have developed interpretable algorithms along with an instructive graphical user interface to enable a worker to deploy the robot to a new task or update a task in less than an hour. Only afterward do we use self-supervised deep learning to augment the program to allow the robot to become more efficient at a task as it does it.
Branden Romero is currently the Chief Engineer at Robotic Materials Inc. He received his BS in Computer Science from the University of Colorado Boulder in 2017, and is going to pursue a Ph.D. in Computer Science from Massachusetts Institute of Technology beginning Fall 2018. He has received three awards from the IEEE/RSJ International Conference of Intelligent Robots and Systems Robotic Grasping and Manipulation Competitions.