|
|
Reasoning under uncertain observations in robots
Jeremy Wyatt
University of Birmingham
Monday 20 June 2011, 1:30-2:30pm,
Foster Court 101, UCL
Reasoning under uncertain observations in robots
Robots act under noise in observation and system dynamics. One formalism for modelling this is the POMDP. In practice, however, POMDP solvers don't scale to the size of problem we typically encounter in robotics. Various approximate problems thus need to be constructed. In this talk I will give an overview of four problems we have solved with different approximations. The first is observation planning during search for hydrothermal vents with an AUV. The second is controlling gaze during a simple manipulation task with a robot. Both of these are tested on simulations. The third problem involves control of visual processing in a static scene when performing question answering. The fourth concerns object search in a home or office environment with a real robot. The tricks we apply include using entropic measures as rewards, using hierarchy in our POMDPs, using one step lookahead, and combining POMDP planning with more conventional planning algorithms. The take home message is that in robotics we have to make judicious choices about how to approximate the right problem in order to achieve reasonable results.
PLEASE NOTE ROOM CHANGE: Foster Court 101, UCL
