Drift-free laser-based Localisation of a Humanoid Robot

Research in the field of humanoid robotics has been pushed forward by the DARPA Robotics Challenge. Humanoid robots were performing complex balancing tasks and manipulating a variety of objects while being supported by the human operator at the level of decision making. Future work aims for achieving task-level autonomy. One of the key requirements for a robot to perform a task autonomously is to have an accurate estimate of its current state (i.e. its pose) in the world. If it is wrong or unsure about its own pose, the robot will also misinterpret the information perceived about its surroundings through its on-board sensors and expect the objects in the world to be in different poses than the actual ones. Typically, the state of the art techniques for state estimation tested on real robots are based on proprioceptive measurements from joint encoders and inertial sensors. However, the estimates computed using only proprioceptive sensing are affected by drift accumulated over time. A correction to this drift can be computed using exteroceptive sensors such as laser scanners etc. In this talk I will focus on drift-free localization for a humanoid robot using a 3D point cloud registration technique. Specifically, I will discuss the challenges that arise due the necessity of aligning several subsequent clouds to the same model as the robot moves around in the environment. Additionally, we evaluate the performance of ICP-based 3D point cloud registration in dynamic semi-structured environments and propose a non-incremental 3D registration strategy to improve the accuracy of the robot's state estimate.