Hierarchical Control of Limbless Locomotion Using a Bio-inspired CPG Model

Abstract

Limbless robots have the potential ability to perform various highly efficient movements in different environments, taking advantage of the features of limbless locomotion, such as a low center of gravity, a large contact area and a distributed mass. This thesis deals with the locomotion control of limbless robots, concentrating specifically on the study of a hierarchical control architecture as steps toward developing limbless robots capable of 3D locomotion, fast reflex responses and sophisticated responses to environmental stimuli.

First, an overview of limbless robots is presented. Various limbless robots found in the literature are investigated. The survey not only introduces some potential applications for limbless robots, but also establishes a classification of limbless locomotion according to the limbless robots\\\' configurations and auxiliary equipment. Moreover, different approaches to autonomously generate motion patterns for limbless robots are discussed. One type of control approaches based on Central Pattern Generators (CPGs) is emphasized, since it is ideally suited to being applied to a hierarchical control architecture.

Then, a bio-inspired CPG model is proposed. The key problem for developing such a hierarchical control architecture is how to design a CPG based controller that can not only generate various gaits, but also provide a solution for realizing reflex mechanisms as well as integrating sensory feedback. To this end, a CPG model inspired by the neuronal circuit diagram in the spinal cord of swimming lampreys is designed. A set of interneurons described with sigmoid functions and leaky integrators is incorporated into the design of the neural oscillator for rhythmic signal generation. Furthermore, according to the connection between neural oscillators, a chained type and a cyclic type of CPG circuits are developed. The chained type CPG circuit is used for generating traveling waves between oscillators, while the cyclic type CPG circuit is used for producing synchronization and maintenance activities. Through numerical simulations, the control parameters over relevant characteristics of the two types of CPG circuits are studied in detail.

Next, the proposed CPG model is further designed for limbless gait implementation. Considering the configuration of limbless robots with pitch modules and yaw modules connected alternatively, two CPG circuits are applied to the pitch grouped modules and the yaw grouped modules, respectively. Both the necessary conditions for cooperation between the two CPG circuits and the control parameters for fast limbless locomotion are investigated. Four types of limbless gaits, i.e. side winding, rolling, turning and flapping are realized. Results of simulations and experiments show the effectiveness of the proposed CPG circuits in generating limbless locomotion.

After that, in order to realize fast sensory reflex responses, the concept of both sensory neurons and reflex arcs are utilized. Since the proposed CPG model is derived from neural circuit in the spinal cord of lampreys and the existence of sensory neurons in lampreys has been proven, it is simple and natural to add sensory neurons into the proposed CPG model at the neuronal level. Based on the design of the sensory neurons, a reflex mechanism taking advantage of reflex arcs forms short pathways to bridge external stimuli and the CPG model. Thus fast responses can be made when the external stimuli are afferent to the CPG model. A ball hitting experiment and a corridor passing experiment confirm the feasibility of the reflex mechanism.

Finally, the development of sophisticated responses to environmental stimuli is presented. A framework that combines the CPG model with a learning method is proposed for achieving adaptive limbless locomotion. The key issue of the framework is to find a mapping that converts external stimuli to proper sensory input, so as to modify the output of the CPG model and thus enable the limbless robot to adapt to environments. Two types of learning methods, i.e. a genetic algorithm (GA) based method and a reinforcement learning (RL) based method are applied to the framework, respectively. Through a slope climbing experiment, it is verified that both of them can achieve adaptive limbless locomotion. Furthermore, the performance of adaptive limbless locomotion under the two methods is compared and analyzed, which provides future work with the possible solutions for promoting the performance of adaptive limbless locomotion.

From the results of simulations and experiments, the hierarchical control architecture is confirmed to be a solid platform for improving the locomotive behaviors of limbless robots.