Tuesday 12pm, 10 February 2015

Screen shot 2015 02 09 at 4.23.15 pm

Multi-Armed Bandits for Robotic Grasping Under Uncertainty

Michael Laskey and Jeff Mahler

PhD Student - UC Berkeley


The multi-armed bandit model is a statistical model of an agent attempting to make a sequence of correct decisions while concurrently gathering information about each possible decision. Solutions to the multi-armed bandit model have been used in applications for which evaluating all possible options is expensive or impossible, such as the optimal design of clinical trials, market pricing, and choosing strategies for games. This talk will explore an interesting application of applying the MAB formulation to the problem of a robot selecting a grasp under uncertainty. Low cost sensors, like the Kinect, lead to uncertainty in object shape and pose, which can make grasp planning computationally expensive. Bandit algorithms can provide an efficient policy for determining a grasp plan with high confidence of success, which we demonstrate with empirical results. Using robotic grasping as a case study, we will describe bandit models, such as Bayesian Algorithms, Best-Arm Identification and Contextual Bandits.


Michael is a Graduate Research Fellow at the University of California, Berkeley, pursing a PhD in Electrical Engineering and Computer Science. He researches the application of sequential decision making applied to robotics in collaboration with the Automation Science Lab. Michael previously worked on fabrication of nano-fluidic devices with femto-second lasers at the University of Michigan, Ann Arbor, where he received a B.S. in Electrical Engineering. His other research interests include human-robot collaboration,belief space planning, reinforcement learning and the application of Bayesian non-parametric techniques for robotics.

Jeff is a Graduate Student Researcher at the University of California, Berkeley, pursing a PhD in Electrical Engineering and Computer Science. His research is on handling uncertainty in robotic perception in control in the Automation Science Lab, particularly in the context of manipulation. He is interested in applications requiring low-cost sensors, such as small-scale manufacturing, and those with sensing constraints, such as in robot-assisted surgery. Jeff graduated with a B.S. in Electrical Engineering from the University of Texas at Austin, where he worked on handheld and low-cost 3D scanning with the Kinect and cofounded the 3D modeling startup Lynx Laboratories.