Simple yet Effective Technique for Robust Real-time Instability Detection for Humanoid Robots Using Minimal Sensor Input

Legged locomotion of autonomous humanoid robots is advantageous but also challenging since it inherently suffers from high posture instability. External disturbances such as collisions with other objects or robots in the environment can cause a robot to fall. Many of the existing approaches for instability detection and falling prevention include a large number of sensors resulting in complex multi-sensor data fusion and are not decoupled from the walking motion planning. Such methods can not simply be integrated into an existing low-level controller for real-time motion generation and stabilization of a humanoid robot. A procedure that is both easily implementable using a minimal number of affordable sensors and capable of reliable detection of posture instabilities is missing to date. We propose a simple, yet reliable balance control technique consisting of a filtering module for the used data from two-axesgyroscopes and -accelerometers located at the trunk, an instability classification algorithm, and a lunge step module. The modules are implemented on our humanoid robots which participate at the yearly RoboCup competitions in the humanoid kid-size league of soccer playing robots. Experimental results show that the approach is suited for real-time operation during walking.