Kinematic Invariant Parametric Snake Robot Gait Design for Enceladus Exploration

Snake-inspired robots demonstrate versatility through challenging terrains. However, their highdimensionalcontinuous action spaces make analytical gait design difficult. Early pioneers showedgait parameterization over low-dimensional spatially and temporally varying sine waves can serveas basis functions in gait shape space. Optimizing such gaits is a non-convex problem as parameterspace is fraught with local optima exacerbated by constraints such as actuator limits. ReinforcementLearning (RL) has emerged as a promising alternative for gait search. However, end-to-endRL approaches lack safety guarantees and don’t yield an intuitive representation of the learnedgait. We propose a hybrid method that first identifies sidewinding gaits for novel bend-twist kinematicchains using RL, before distilling the gait into an equivalent open-loop parametric representation.Our method avoids the twist-windup problem (a key challenge we identified with priorwork) while combining task-space learning with interpretable policy approximation. Simulationand hardware experiments demonstrate our proposed method generates parametric gaits for a simpleline-following task where an existing optimization-based curve-fitting method cannot.Keywords: reinforcement learning, snake robotics, gait design, parametric curve fitting.