Warning:
This page is
provided for historical and archival purposes
only. While the seminar dates are correct, we offer no
guarantee of informational accuracy or link
validity. Contact information for the speakers, hosts and
seminar committee are certainly out of date.
Assistant Professor
The Robotics Institute and School of Computer Science
Carnegie Mellon University
The goal of my research is to understand the relationship between sensing, action, and prediction. I have explored various extreme points in this space. With Matt Mason I looked at sensorless strategies, for my thesis work I looked at randomized strategies, and currently I am investigating fast-action minimal-sensing strategies. My goal is to understand precisely what sensing is required to accomplish a manipulation task quickly, what the tradeoffs are between sensing, action, and prediction, and how to design optimal sensors and error-tolerant strategies.
This work is motivated by several desires. First, I would like to program robots more easily than is currently possible. Second, I would like to understand the scope and limitations of autonomous systems, whether biological or artificial. Third, I would like to reduce the complexity of design and planning by codifying the design parameters required to achieve a given level of automation.
At a more detailed level, my current research investigates a method for automatically designing sensors from the specification of a robot's task, its repertoire of actions, and its uncertainty in control. This method creates abstract sensors whose purpose it is to provide precisely the information required by the robot. The approach is to generate an action-based plan for accomplishing the task, initially under the assumption of perfect sensing. By examining the structure of this plan, one can then often reduce the actual sensing required from perfect sensing to some smaller more tractable amount of sensing. In short, the plan specifies a sensor that delivers the information required for the plan to function correctly.
We have examined this approach in the context of several simple manipulation tasks. These include, of course, peg-in-hole, but also some extensions such as the two-pin-two-hole, screw-plate-table, and balancing tasks. In addition, we have run into some theoretical discoveries. The main discovery is that any strategy for accomplishing a task must in the worst case establish the same information as would be provided by the abstract sensor discussed above. This suggests that we view automated robot programming as a method for augmenting or perturbing the basic plan created with a perfect sensor and imperfect control. The type of augmentation depends on the physical sensor available and the amount of prediction one is willing to compute. Much of this work remains tantalizingly open.
Host: Yangsheng Xu (xu@cs.cmu.edu) Appointment: Lalit Katragadda (lalit@cs.cmu.edu)