Event Location: NSH 3305

Abstract: Consider a robot arm with a hand-mounted sensor. In order to interact with and understand the world, the robot must be able to reconstruct it using its sensor by moving its joints to scan the scene.

If the robot’s kinematics are known with absolute certainty, the problem is the simple 3D mapping problem. However, robot arms often suffer from actuator noise caused by unknown dynamics (such as cable stretch, springs, deformable links). Given such noise, getting a good 3D reconstruction is often impossible without some form of robot pose estimation.

In this work, I propose a novel solution to the problem of 3D reconstruction using robot arms combining ideas from the manipulator pose estimation and visual SLAM literature. My main insight is that by solving the simultaneous localization and mapping (SLAM) problem in the configuration space of the robot, rather than in the 6DOF pose space of the camera, the quality of 3D reconstruction can be dramatically improved — and by using the configuration of the robot as the state space of the problem, we are able to localize the robot, control it, and reconstruct the scene simultaneously.

Committee:Sidd Sirinivasa, Co-chair

Michael Kaess, Co-chair

George Kantor

Andrew Davison, Imperial College London