Abstract:

Perception on humanoid robots presents several unique challenges. Approaches to localization and mapping must deliver
accurate results to comply with the small error tolerances imposed by the walking controller if the robot is to successfully
step onto surfaces or avoid obstacles. Moreover, they must be able to deal with rapid scene changes, large camera
displacement and camera shakiness and should operate in real-time, since pausing for deliberation or sensing is often
not an option. However, the complexity of vision processing often implies that these requirements cannot all be met at
once with the traditional CPU-based computational resources available. In this paper, we present a GPU implementation of
a model-based 3D tracking algorithm which we have applied specifically to the problem of humanoid locomotion. Our
system robustly fits the visible model edges of a given object to edge features extracted from the video stream, yielding
the full 6DOF pose of the object relative to the camera. The recovered pose, together with the robot kinematics, allows
us to accurately localize the robot with respect to the object and to generate environment maps. These can then be used to
plan a sequence of footsteps that, when executed, allow the robot to quickly and successfully circumnavigate obstacles
and climb stairs.