Localisation for Autonomous Humanoid Navigation

Abstract

Autonomous humanoid navigation in non-trivial environments requires high precision accuracy due to the difficulty in achieving stable bipedal locomotion. In particular, an accurate localisation estimate is needed to plan footstep placement on a narrow staircase. This paper reports the development of an accurate 6DOF particle filter based localisation system for a humanoid robot moving within a known 2.5 dimensional map. A laser range sensor mounted within the robot's head makes 120 degree planar scans of the environment up to a distance of 4 meters. Localisation accuracy is achieved by carefully characterising the robot's odometry model, introducing a novel motion model for particle prediction and decoupling the bounded and unbounded components of humanoid position uncertainty. The novel motion predicts a more accurate particle distribution by modeling the motion of a humanoid robot and including uncertainty in sampling time as well as position accuracy. In addition the reported system estimates the localisation uncertainty distribution and implements a model based gaze attraction behaviour to further reduce localisation error.