Visual-inertial teach and repeat

Abstract

Teach and Repeat (T&R) refers to the technology that allows a robot to autonomously follow a previously traversed route, in a natural scene and using only its onboard sensors. In this paper we present a Visual-Inertial Teach and Repeat (VI-T&R) algorithm that uses stereo and inertial data and targets Unmanned Aerial Vehicles with limited on-board computational resources. We propose a tightly-coupled relative formulation of the visual-inertial constraints that is tailored to the T&R application. In order to achieve real-time operation on limited hardware, we reduce the problem to motion-only visual-inertial Bundle Adjustment. In the repeat stage, we detail how to generate a trajectory and smoothly follow it with a constantly changing relative frame. The proposed method is validated in simulated environments, using real sensor data from the public EuRoC dataset, and using our own robotic setup and closed-loop control. Our experimental results demonstrate high accuracy and real-time performance both on a standard desktop system and on a low-cost Odroid X-U4 embedded computer.