Inertial-Joint Learning-aided Robust Pose Estimation for Legged Robots

Abstract

Legged robots have received much attention recently due to their maneuverability in various environments. For autonomously operating the legged robots in harsh conditions like bushes, mountains, and mud, stable and accurate state estimation is necessary. However, estimating the robot’s pose under such slippery and visually not informative conditions is hard. Therefore, this paper proposes a novel framework for the legged robot’s state estimation. Our proposed algorithm utilizes the deep inertial-joint factor based on the inertial-joint network. With real-world experiments, our proposed system validates that the deep inertial-joint factor can help improve the performance of the state estimator.