A Robust State Estimator for Legged Robots using Visual-Inertial-Leg Kinematics in the Wild

Abstract

Legged robots have drawn attention recently because they can overcome obstacles in the wild, and many robust control methods for them have been developed. On the other hand, a precise and robust state estimator for them is relatively underdeveloped. For instance, many estimators only exploit proprioceptive sensors such as an IMU and joint encoders, which accumulate the drift due to slippage. Moreover, they rely on non-slip assumptions invalid in many outdoor environments. Several researchers tried to mitigate this issue by using exteroceptive sensors. However, as the estimator still depends on the non-slip assumption, leg kinematics information is not exploited properly. In this work, we propose how to use leg kinematics information in the wild without any strong assumptions. Specifically, we track the foot velocity coupled with the robot’s state with leg kinematics. In addition, the joint encoder measurements are dealt with preintegration fashion to reduce computational complexity.