Enhancing Proprioceptive Reinforcement Learning-Based Quadrupedal Locomotion Controller via Explicit-Implicit Estimation Network

Abstract

Industrial inspection often involves safety risks, making legged robots highly suitable for replacing human labor. However, robust locomotion across diverse, unstructured terrains remains challenging due to estimation uncertainties. In this work, we propose a proprioceptive reinforcement learning (RL)-based locomotion framework enhanced by a novel explicitimplicit estimation network (EI-Net). The explicit network accurately estimates physically meaningful states, while the implicit network leverages a self-supervised learning approach to decorrelate features, achieving superior terrain categorization. Together, this architecture significantly improves dynamic stability. Simulation and real-world experiments show that the robot traverses stairs up to 60% of its body height and single steps up to 0.22 m, achieving an 80% success rate, outperforming other methods. Notably, the framework achieves zero-shot sim-to-real transfer and strong robustness, completing a 2.25km outdoor traversal with a 3 kg payload across unmodeled terrains.