Learning to calibrate: Reinforcement learning for guided calibration of visual–inertial rigs

Abstract

We present a new approach to assisted intrinsic and extrinsic calibration with an observability-aware visual–inertial calibration system that guides the user through the calibration procedure by suggesting easy-to-perform motions that render the calibration parameters observable. This is done by identifying which subset of the parameter space is rendered observable with a rank-revealing decomposition of the Fisher information matrix, modeling calibration as a Markov decision process and using reinforcement learning to establish which discrete sequence of motions optimizes for the regression of the desired parameters. The goal is to address the assumption common to most calibration solutions: that sufficiently informative motions are provided by the operator. We do not make use of a process model and instead leverage an experience-based approach that is broadly applicable to any platform in the context of simultaneous localization and mapping. This is a step in the direction of long-term autonomy and “power-on-and-go” robotic systems, making repeatable and reliable calibration accessible to the non-expert operator.

Calib-Net: Calibrating the Low-Cost IMU via Deep Convolutional Neural Network

“…As to the usage of deep learning in IMU calibration and propagation, Yan et al (2018) proposed to use the machine learning technology to regress velocity vector with linear accelerations and angular velocities as inputs. Nobre and Heckman (2019) proposed to model the IMU calibration as a Markov Decision Process (MDP) and use reinforcement learning to achieve the regression of calibration parameters. Clark et al (2017) presented VINet which takes the visual-inertial odometry problem as a sequence-to-sequence learning problem to solve and avoids manual camera/IMU calibration operation.…”

Unified Data Collection for Visual-Inertial Calibration via Deep Reinforcement Learning

“…RL based methods can better include those general requirements. Nobre et al [9] defined a library containing different motions and applied Q-learning to select a sequence of motions from the library with the goal to obtain enough observability for calibration. However, the performance, as well as efficiency of calibration, is constrained by the predefined library because many possible motion sequences are not included or explored.…”

“…Precise calibration, which for VI sensors refers to the parameters for the camera intrinsics, the camera-IMU extrinsics, and the time offset between the different sensors, are of great importance to the accuracy and performance of VI systems [6], [7], [8]. However, it is usually non-trivial to perform the calibration by hand or with a manually pre-programmed operator, since it often requires complex motion routines ensuring observability in controlled environments [8], [9], [10], [11]. In order to guarantee high calibration accuracy, typically large amounts of data are recorded for optimization.…”

“…In early studies, several works applied reinforcement learning (RL) to get the best sequence of trajectories for VI calibration. Nobre et al [9] proposed using Q-learning to select a sequence of motions from a predefined library aiming to render sufficient observability of the calibration problem. However, this approach only suggests which predefined motions to choose from the empirically designed library without exploring new possible motion primitives.…”

Towards Online Observability-Aware Trajectory Optimization for Landmark-based Estimators

“…For example, [1,3,11] identify and enumerate non-observable trajectories to hand-design a discrete set of "wellobservable" maneuvers for use in online planning. More recently, [14] leverages a reinforcement learning framework to select a sequence of primitive maneuvers in a manual calibration routine. From the side of continuous optimization, [10,15] maximize observability metrics based on the Local Observability Gramian (LOG) in a continuous-optimization setting.…”