Measuring impacts using Support Vector Machines on a standing humanoid robot

The proprioceptive sense provides a wealth of information to the human brain to be used for all forms of motion control and tactile perception. A humanoid robot is also equipped with a proprioceptive sense, and is typically one of its most accurate. However, proprioception is under utilised in humanoid robots, outside of simple joint trajectory tracking, and can be used to improve the control of stance. In this paper we develop a system that makes use of the proprioceptive sense, in particular the joint velocities, to perceive and quantify external perturbations to a standing humanoid robot. An optimised threshold detector is used to perceive perturbations with an average detection delay of 138 ms. A Support Vector Machine is utilised to localise the contact point of the perturbation to one of 16 discretised locations on the upper and lower body with 100% success rate.The direction and strength of the external perturbation is then estimated using a pair of orthogonal Support Vector Regression models. When the two models are applied to an omnidirectional perturbation-set the direction and strength are predicted with a mean-squared-error of 4.2° and 0.087Ns, respectively. The system provides useful information that could be used to improve the quality of a corrective response to a perturbation. Additionally, the system provides a sense of `touch´ without the need for a tactile skin.