Environmental contact modeling for the earthworm-like robot via the novel elementary mechanical network

• Published in: Science China. Technological sciences Vol. 65; no. 6; pp. 1366 - 1382
• Main Authors: He, KunPeng, Zhang, XiaoXu
• Format: Journal Article
• Language: English
• Published: Beijing Science China Press 01.06.2022; Springer Nature B.V
• Subjects: Boolean algebra; Coulomb friction; Empirical analysis; Engineering; Environment models; Interaction models; Locomotion; Modules; Particle swarm optimization; Reconfiguration; Recurrent neural networks; Reliability analysis; Robot dynamics; Robots; Sand; Weightlifting; Worms; reconfigurable network; bioinspired locomotion robot; system identification; soft terrain; data-driven modeling
• ISSN: 1674-7321; 1869-1900
• DOI: 10.1007/s11431-021-2016-8

The earthworm-like robot is designed for prospective applications such as disaster rescue and pipeline detection in natural environments. However, current studies on the interaction modeling between the earthworm-like robot and the environment only consider rigid contact. This simplification limits the reliability of dynamic analysis and locomotion optimization on soft surfaces, such as sand. Therefore, developing a method for refined contact modeling for the earthworm-like robot and describing the contact effect induced by the soft environmental medium is urgent. To this end, this paper proposes a new modeling architecture called the elementary mechanical network (EMN). EMN is constructed as fractal structures for the convenience of network reconfiguration. First, elementary mechanical elements, such as the damper, spring, and slider, are parallelly connected to constitute a basic module. Second, the modules are serially linked to create a group. Finally, the groups are parallelly assembled to form the network. EMN is also proven to be equivalent to recurrent neural networks in specific forms. Therefore, EMN inherits the advantages of physical interpretability from mechanical elements and universal approximability from conventional networks. In addition, particle swarm optimization and Boolean operation are employed for network weight training and topological minimization. Numerical examples show that using EMNs with identical initial structures can accurately describe diverse empirical models in the minimum realization. EMN is applied for contact modeling for the earthworm-like locomotion robot in the dry sand based on such versatility. The experiment measures the normal and tangential ground reaction forces with different sinkage depths and locomotion speeds. Trained results reveal a common law that the contact effect in the dry sand is similar to Coulomb friction. The proposed EMN does not require prior system knowledge and promises a minimal physical representation, thus encouraging a successful exploration of constitutive modeling in broad scopes.