Computer Science Graduate Seminar: Uncertainty-Aware Perceptual Robot Learning
Thursday, June 27, 2019, 2:15pm
Location: UMIC Research Centre, room 025, Mies-van-der-Rohe-Straße 15
Speaker: Prof. Dr. David Held, The Robotics Institute, Carnegie Mellon University
Robot learning systems need to deal with perceptual uncertainty, due to a variety of factors such as occlusions, sensor noise, small objects, limited capacity models, and novel objects. We believe that perceptual systems should be aware of their uncertainty and that decision-making algorithms should incorporate such uncertainty.
We present three current directions for achieving this.
First, we present an object instance detection system that combines machine learning with correspondence matching to verify the proposed detections and reject uncertain detections.
Next, we show an approach for estimate a distribution over orientation uncertainty by augmenting any deep pose estimation system.
Last, we present an approach for reinforcement learning with deformable objects that does not assume access to the ground-truth state.
Together, these directions represent our current approach to dealing with perceptual uncertainty in robot learning systems.
The computer science lecturers invite interested people to join.