Learning Slip Detection for Agile Locomotion of Quadruped Robots

Abstract

The very often occurrence of foot slippage imposes a significant challenge on the locomotion of multi-legged robots, and slip detection is the prerequisite to prevent robots from slippage. Slip detection methods based on additive sensors are susceptible to noise and the sensors are expensive and easily damaged, and the existing methods using proprioceptive information fail to distinguish which feet are slipping given the robot’s state in case multiple legs slip at the same time, due to strong inter-dependence among the body and multiple branches of leg mechanisms. Deep learning can establish the relationship between input and output using a large number of samples, avoiding the complex process of manual feature extraction and giving hope for a solution to this difficulty. In this paper, we present a deep learning-based approach to slip detection in quadruped robots. Our slip detector uses a convolutional neural network to fuse information acquired from proprioceptive sensors to infer the slip state to provide a priori knowledge to the controller. Simulation results indicate the effectiveness of the proposed approach in achieving accurate and immediate slip detection of each foot for agile locomotion without prior knowledge about the contact state in multiple scenarios, and the learning accuracy is greater than 95% in every scenario.