VIR-SLAM: Visual, Inertial, and Ranging SLAM for single and multi-robot systems

Yanjun Cao, Giovanni BeltramearXiv 2020 · arXiv

In the world of autonomous robotics, Visual-Inertial Odometry (VIO) is a gold standard for local navigation. By combining a monocular camera with an Inertial Measurement Unit (IMU), robots can estimate their movement with impressive precision. However, VIO has a fatal flaw: drift. Over long trajectories or in visually repetitive environments (like a long, white hallway), small errors accumulate, causing the robot’s estimated path to veer away from reality.

VIR-SLAM (Visual, Inertial, and Ranging SLAM) offers a clever solution. By adding just one low-cost Ultra-Wideband (UWB) anchor to the environment, researchers Yanjun Cao and Giovanni Beltrame have developed a system that significantly suppresses drift and simplifies collaborative multi-robot mapping.

TL;DR

VIR-SLAM enhances standard visual-inertial navigation by integrating distance measurements from a single, static UWB anchor. This approach reduces localization error by over 20% compared to state-of-the-art VIO and enables multiple robots to merge their maps efficiently without needing to "see" the same visual landmarks.

What Problem Does This Paper Address?

While traditional SLAM (Simultaneous Localization and Mapping) uses "loop closures"—recognizing a previously visited place to reset accumulated error—this relies heavily on the environment. If the lighting changes or the area is featureless (like a glass-walled atrium), visual loop closure fails.

Furthermore, in multi-robot systems, aligning the maps of two different robots usually requires them to observe the same visual features simultaneously or exchange massive feature databases. VIR-SLAM aims to solve these issues using UWB ranging, which provides distance data that is immune to lighting conditions and carries built-in identification of the signal source.

Methods

The core innovation of VIR-SLAM lies in its double-layer sliding window optimization.

  1. Short Window (Local): This layer handles high-frequency visual and IMU data, similar to standard VIO. It provides accurate short-term relative pose estimation.
  2. Long Window (Global): This layer maintains a history of robot poses specifically associated with UWB range measurements to the static anchor. By keeping these distant poses in the optimization, the system can use the absolute distance to the anchor to "pull" the trajectory back into alignment, correcting long-term drift.
Factor graph showing the double-layer window structure. The long window (orange) incorporates UWB factors, while the short window (gray) handles visual and IMU factors.

For multi-robot scenarios, the authors leverage the fact that all robots can determine the direction of gravity using their IMUs. This means they only need to find the relative yaw (rotation) and 2D translation between their coordinate frames. By sharing their estimated anchor positions and taking just two mutual range measurements, robots can calculate a transformation matrix to fuse their maps at almost no extra computational cost.

Datasets

The authors validated VIR-SLAM through three primary avenues:

  • EuRoC MAV Dataset: A standard benchmark for VIO. The authors simulated UWB ranging by adding Gaussian noise to the ground truth data.
  • Controlled Lab Experiments: Using Spiri robots equipped with Realsense cameras and UWB modules, with an OptiTrack motion capture system providing ground truth.
  • Challenging Atrium Test: A real-world test in a large atrium with featureless walls and glass partitions—a "nightmare" scenario for pure visual SLAM.

Key Insights and Ideas

  • Infrastructure Minimalism: Unlike most UWB systems that require 3 or 4 anchors for trilateration (like a local GPS), VIR-SLAM proves that one anchor is enough when fused with VIO. This makes it practical for rapid deployment in unknown environments.
  • UWB as a Communication Channel: UWB isn't just for ranging; it can carry small data packets. The authors suggest that pose estimates can be swapped during the ranging process, making multi-robot coordination extremely low-bandwidth.
  • Decoupling Drift from Vision: By using a "long window" for UWB, the system corrects drift even when the camera is struggling with poor features, effectively decoupling global consistency from visual quality.

Key Results

  • Accuracy Boost: On the EuRoC dataset, VIR-SLAM outperformed VINS-Mono (a state-of-the-art VIO) by an average of 20%, with improvements as high as 24.8% in some sequences.
  • Robustness in "Blind" Environments: In the atrium test, standard VIO drifted by over 5.4 meters. VIR-SLAM maintained a 2D error of just 0.148 meters.
  • Multi-Robot Success: The system successfully merged trajectories from two independent robots to form a combined map (drawing the "MIST" lab logo) using only two inter-robot measurements.
Multi-robot experiment: Robot 1 (orange) and Robot 2 (green) merge their independent trajectories into a single coordinate frame to form the "MIST" logo.

Limitations and Future Work

The primary limitation noted is the z-axis (vertical) error. If the UWB anchor is placed at roughly the same height as the robot's flight path, the ranging data provides very little information to correct vertical drift. The authors observed higher 3D errors (0.853m) compared to 2D errors (0.148m) in these cases.

Future work could involve strategies for "dropping" anchors dynamically during exploration or investigating how multiple anchors (when available) could further harden the system against z-axis drift.

Reference: Cao, Y., & Beltrame, G. (2020). VIR-SLAM: Visual, Inertial, and Ranging SLAM for single and multi-robot systems. arXiv:2006.00420v1.

Summary updated

Abstract

Monocular cameras coupled with inertial measurements generally give high performance visual inertial odometry. However, drift can be significant with long trajectories, especially when the environment is visually challenging. In this paper, we propose a system that leverages ultra-wideband ranging with one static anchor placed in the environment to correct the accumulated error whenever the anchor is visible. We also use this setup for collaborative SLAM: different robots use mutual ranging (when available) and the common anchor to estimate the transformation between each other, facilitating map fusion Our system consists of two modules: a double layer ranging, visual, and inertial odometry for single robots, and a transformation estimation module for collaborative SLAM. We test our system on public datasets by simulating an ultra-wideband sensor as well as on real robots. Experiments show our method can outperform state-of-the-art visual-inertial odometry by more than 20%. For visually challenging environments, our method works even the visual-inertial odometry has significant drift Furthermore, we can compute the collaborative SLAM transformation matrix at almost no extra computation cost.
BibTeX
@misc{cao2020vir,
  author = {Yanjun Cao and Giovanni Beltrame},
  title = {{VIR-SLAM: Visual, Inertial, and Ranging SLAM for single and multi-robot systems}},
  year = {2020},
  eprint = {2006.00420},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2006.00420},
}

Loading related content...

Figure