Fast Haptic Terrain Classification for Legged Robots Using Transformer

Abstract

The haptic terrain classification is an essential component of a mobile walking robot control system, ensuring proper gait adaptation to the changing environmental conditions. In this work, we further tackle this problem with force and torque measurements from feet while focusing on real-life applicability defined as low computational demand and rapid inference time. To meet these requirements, we propose two classical machine learning algorithms (DTW-KNN and ROCKET) and two deep-learning solutions – a typical feed-forward solution based on temporal convolution network (TCN) and the currently prevailing transformer architecture. The experiments conducted on the publicly available haptic classification dataset revealed that we could reach classification results marginally lower than state of the art with networks containing up to 50 times fewer parameters within an improved inference time of several milliseconds on a CPU.