Force-Sensor-Based Walking-Environment Recognition of Biped Robots

Usability of biped robots in real applications depends on the robot capability of stable and robust walking with high efficiency. To cope with the various practical challenges, the robot must therefore be able to recognize its environment properties. We report a system for indoor-surface detection based on force sensors attached below the feet of the robot. To verify the recognition performance, indoor evaluation surfaces with 5 different properties are used. The capability of fast surface-property recognition is realized by processing the stream of force-sensor data according to the method of overlapping sliding windows, in order to generate 4 different features in a dynamic way. A k-nearest-neighbor (kNN) classifier with multiple classes is applied for real-time high- accuracy recognition of the surface-specific robot-walking characteristics. In particular, recognition performance can be increased by combining the studied features into a single feature descriptor, instead of using each feature separately. Achievability of an overall accuracy of 90.4% and an average precision of 91.49% is verified. Thus, a favorable trade-off between cost and performance is realized. The developed method is useful for optimized dynamic robot-body balancing and walking-speed adjustment, according to the recognized surface properties.