[Paper Review] MINS: Efficient and Robust Multisensor-aided Inertial Navigation System

Woosik Lee, Patrick Geneva+2 more|arXiv (Cornell University)|Sep 27, 2023

Robotics and Sensor-Based LocalizationEngineering3 citations

TL;DR

This paper proposes MINS, a tightly-coupled, efficient, and robust multisensor-aided inertial navigation system that fuses IMUs, wheel encoders, cameras, LiDARs, and GNSS in a continuous-time filtering framework. It introduces high-order on-manifold interpolation, dynamic cloning for adaptive computational complexity, online spatiotemporal calibration, and IMU-wheel-only initialization, achieving state-of-the-art accuracy, consistency, and real-time performance with minimal computation, even under sensor failures or dynamic conditions.

ABSTRACT

Robust multisensor fusion of multi-modal measurements such as IMUs, wheel encoders, cameras, LiDARs, and GPS holds great potential due to its innate ability to improve resilience to sensor failures and measurement outliers, thereby enabling robust autonomy. To the best of our knowledge, this work is among the first to develop a consistent tightly-coupled Multisensor-aided Inertial Navigation System (MINS) that is capable of fusing the most common navigation sensors in an efficient filtering framework, by addressing the particular challenges of computational complexity, sensor asynchronicity, and intra-sensor calibration. In particular, we propose a consistent high-order on-manifold interpolation scheme to enable efficient asynchronous sensor fusion and state management strategy (i.e. dynamic cloning). The proposed dynamic cloning leverages motion-induced information to adaptively select interpolation orders to control computational complexity while minimizing trajectory representation errors. We perform online intrinsic and extrinsic (spatiotemporal) calibration of all onboard sensors to compensate for poor prior calibration and/or degraded calibration varying over time. Additionally, we develop an initialization method with only proprioceptive measurements of IMU and wheel encoders, instead of exteroceptive sensors, which is shown to be less affected by the environment and more robust in highly dynamic scenarios. We extensively validate the proposed MINS in simulations and large-scale challenging real-world datasets, outperforming the existing state-of-the-art methods, in terms of localization accuracy, consistency, and computation efficiency. We have also open-sourced our algorithm, simulator, and evaluation toolbox for the benefit of the community: https://github.com/rpng/mins.

Motivation & Objective

• To address the lack of efficient, tightly-coupled multisensor fusion systems that integrate five common navigation sensors (IMU, wheel encoders, cameras, LiDAR, GNSS) in real time.
• To overcome challenges in asynchronous sensor data, high computational complexity, and time-varying sensor calibration in long-term operation.
• To develop a robust initialization method that relies only on proprioceptive sensors (IMU and wheel encoders) for high-dynamic or exteroceptive-sensor-degraded environments.
• To enable consistent, high-order on-manifold state interpolation that minimizes trajectory representation errors while maintaining computational efficiency.
• To provide online intrinsic and extrinsic (spatiotemporal) calibration to maintain accuracy despite poor or drifting prior calibrations.

Proposed method

• Proposes a continuous-time state representation to reduce computational complexity for asynchronous and delayed measurements.
• Introduces a high-order on-manifold interpolation scheme that ensures consistency and accuracy in state estimation across asynchronous sensor updates.
• Develops a dynamic cloning algorithm that adaptively adjusts interpolation order based on motion dynamics to balance computational load and trajectory fidelity.
• Employs online intrinsic and extrinsic calibration for all sensors to correct for time-varying biases and misalignments without relying on prior calibration.
• Designs a novel initialization technique using only IMU and wheel encoder data, enabling robust start-up in dynamic or exteroceptive-sensor-challenged environments.
• Uses measurement compression and a small state size in the EKF update to maintain low computational overhead, achieving only 1 ms per EKF update regardless of sensor count.

Experimental results

Research questions

• RQ1 Can a tightly-coupled multisensor fusion system efficiently and robustly integrate five common navigation sensors (IMU, wheel encoders, cameras, LiDAR, GNSS) in real time?
• RQ2 How can trajectory representation errors from continuous-time state interpolation be minimized while maintaining computational efficiency?
• RQ3 Can online spatiotemporal calibration improve robustness to poor or drifting sensor calibration in long-term autonomous operation?
• RQ4 Is it possible to achieve reliable system initialization using only proprioceptive sensors (IMU and wheel encoders) in highly dynamic or exteroceptive-sensor-degraded scenarios?
• RQ5 To what extent does dynamic cloning reduce computational complexity without sacrificing estimation accuracy or consistency?

Key findings

• MINS achieves a 1 ms EKF update time regardless of the number of sensors, demonstrating exceptional computational efficiency.
• The system maintains consistent performance across diverse environments, including GNSS-denied, LiDAR-degraded, and camera-occluded scenarios, due to sensor redundancy and robust fusion.
• The dynamic cloning algorithm effectively reduces computational load while preserving trajectory accuracy, outperforming fixed-order interpolation in fidelity and efficiency.
• Online calibration significantly improves estimation consistency and accuracy, especially under time-varying sensor biases or poor initial calibration.
• The IMU-wheel-only initialization method enables successful system startup in highly dynamic scenes where exteroceptive sensors fail, outperforming conventional methods in such conditions.
• In benchmarking on the KAIST Urban 38 dataset, MINS recorded a total computation time of 54–47 ms across different sensor configurations, ranking second-fastest among all evaluated methods, with minimal variation across sensor counts.