Using Elementary Mechanical Networks to Model and Recognize Multi-environment Contacts for the Earthworm-Like Robot

The earthworm-like robot is developed rapidly due to the need for applications such as rescuing in ruins and cleaning pipelines. However, making the robot adaptive to multiple environments is still challenging because the contact effect between the robot and the ground needs to be clarified. To this end, this paper employs a physically interpretable elementary mechanical network (EMN) to mimic the contact effect and then uses the built model to recognize different contact environments. First, training and validation data are collected from experiments that measure the robot’s locomotion and ground reaction force in different environments, e.g., sand, turf, sponge, and acrylic sheet. Then, EMN is trained by merging the particle swarm optimization and Boolean operation, specifically, with identical initial structures for different environment datasets. Finally, based on the contact models represented in EMN, an environment library is established so that machine learning methods can be conveniently implemented to distinguish and recognize the environment where the earthworm-like robot locates. At the conclusion of our study, we achieve a multi-environment recognition accuracy of over 70%. Our findings highlight the potential of EMN models to reveal the nature of multi-environment contacts and enable accurate recognition of the surrounding environment for the earthworm-like robot.