PaperlayerView exact revision on arXiv

Source-linked AI summary

DOOR-SLAM: Distributed, Online, and Outlier Resilient SLAM for Robotic Teams

Pierre-Yves Lajoie, Benjamin Ramtoula, Yun Chang, Luca Carlone, Giovanni Beltrame

arXiv:1909.12198v2cs.RO

TL;DR

Distributed multi-robot SLAM must handle limited communication and perceptual outliers without relying on external localization infrastructure. DOOR-SLAM combines a peer-to-peer, data-efficient front-end with distributed PCM-based outlier rejection and pose graph optimization. Across simulations, datasets, and field tests, it rejects spurious measurements and produces accurate trajectories while supporting less conservative front-end parameters and low communication bandwidth.

  • Problem

    Distributed SLAM is challenged by limited communication and perceptual aliasing, while existing systems are vulnerable to outliers and therefore use conservative loop-closure parameters.

  • Method

    DOOR-SLAM combines peer-to-peer loop-closure detection without raw-data exchange with a distributed pose graph optimizer using PCM to reject spurious measurements.

  • Results

    DOOR-SLAM rejects outliers and produces accurate trajectory estimates across simulations, datasets, and field experiments, including GPS-denied subterranean environments.

  • Takeaways & Limitations

    The robust back-end supports less conservative front-end parameters while the distributed implementation maintains low communication requirements.

Abstract

from arXiv · show

To achieve collaborative tasks, robots in a team need to have a shared understanding of the environment and their location within it. Distributed Simultaneous Localization and Mapping (SLAM) offers a practical solution to localize the robots without relying on an external positioning system (e.g. GPS) and with minimal information exchange. Unfortunately, current distributed SLAM systems are vulnerable to perception outliers and therefore tend to use very conservative parameters for inter-robot place recognition. However, being too conservative comes at the cost of rejecting many valid loop closure candidates, which results in less accurate trajectory estimates. This paper introduces DOOR-SLAM, a fully distributed SLAM system with an outlier rejection mechanism that can work with less conservative parameters. DOOR-SLAM is based on peer-to-peer communication and does not require full connectivity among the robots. DOOR-SLAM includes two key modules: a pose graph optimizer combined with a distributed pairwise consistent measurement set maximization algorithm to reject spurious inter-robot loop closures; and a distributed SLAM front-end that detects inter-robot loop closures without exchanging raw sensor data. The system has been evaluated in simulations, benchmarking datasets, and field experiments, including tests in GPS-denied subterranean environments. DOOR-SLAM produces more inter-robot loop closures, successfully rejects outliers, and results in accurate trajectory estimates, while requiring low communication bandwidth. Full source code is available at https://github.com/MISTLab/DOOR-SLAM.git.

I. INTRODUCTION

DOOR-SLAM addresses distributed multi-robot SLAM challenges caused by limited communication, perceptual aliasing, and outlier loop closures. It combines distributed robust optimization and data-efficient loop-closure detection, and is evaluated across simulations, datasets, and field experiments.

  • External localization infrastructure can be dangerous, expensive, or impractical in settings such as cave exploration and building inspection.
  • Limited bandwidth and communication range make exchanging raw sensor data difficult, while existing systems often assume centralized processing or continuous connectivity.
  • Perceptual aliasing can generate incorrect loop closures between visually similar but distinct places, and distributed outlier rejection methods were lacking.
  • DOOR-SLAM provides peer-to-peer operation without full connectivity, detects inter-robot loop closures without raw-data exchange, rejects outliers, and performs distributed pose graph optimization.
  • Its robust optimizer enables less conservative front-end tuning, while the front-end trades some data efficiency to avoid full connectivity and environment-specific pre-training.
  • The system runs online on NVIDIA Jetson TX2 computers and was evaluated in simulations, KITTI datasets, and GPS-denied subterranean field experiments.

II. RELATED WORK

Related work develops centralized and distributed pose graph optimization and several outlier-mitigation techniques. Distributed approaches reduce reliance on centralized computation, but the literature reviewed here motivates further distributed robustness.

  • Centralized multi-robot pose graph optimization collects all measurements at a central station, whereas distributed methods use local computation and communication.
  • Distributed PGO methods include Jacobi, Gaussian elimination, and Distributed Gauss-Seidel approaches for estimating robot poses.
  • Distributed Gauss-Seidel supports 3D cases and avoids the complex bookkeeping and information double counting required by earlier techniques.
  • Prior outlier-mitigation methods include RANSAC, branch & bound, M-estimation, and binary-variable deactivation approaches.

C. Distributed Loop Closure Detection

Distributed loop-closure detection must identify matches across robots despite bandwidth and communication-range constraints. DOOR-SLAM uses peer-to-peer exchanges within communication range while supporting operation when robots are disconnected.

  • Inter-robot loop closures align robot trajectories in a common reference frame and improve trajectory estimates.
  • Distributed detection is difficult because images are collected at different locations and exchanging them is constrained by communication range and bandwidth.
  • Earlier distributed place-recognition methods reduce bandwidth or scale across teams, but some assume fully connected robots.
  • DOOR-SLAM uses peer-to-peer communication, performing single-robot SLAM when teammates are out of range and distributed SLAM during rendezvous.
  • The system uses onboard stereo imagery, stereo visual odometry, distributed loop-closure detection, and distributed outlier rejection and optimization.

A. Distributed Loop Closure Detection

DOOR-SLAM detects inter-robot loop closures through compact visual descriptors and geometric verification, then rejects inconsistent measurements before distributed pose graph optimization. Pairwise consistency uses existing loop-closure and odometric information, avoiding extra communication.

  • A. Distributed Loop Closure Detection: The loop-closure front-end combines compact-descriptor place recognition with geometric verification of relative poses between robots observing the same scene.
  • A. Distributed Loop Closure Detection: NetVLAD descriptors are computed locally and exchanged when robots enter communication range, with only newly generated descriptors sent after previous encounters.
  • A. Distributed Loop Closure Detection: Geometric verification uses exchanged visual features, descriptors, and 3D positions with solvePnpRansac to obtain inliers and relative pose measurements.
  • B. Distributed Robust PGO: The back-end represents each robot trajectory as keyframe poses and uses odometry plus inter-robot relative-pose measurements.
  • B. Distributed Robust PGO: PCM rejects spurious loop closures caused by perceptual aliasing, which otherwise can produce large trajectory distortions.
  • B. Distributed Robust PGO: PCM checks pairwise consistency by testing whether loop-closure cycles compose to the identity under their noise covariance.
  • B. Distributed Robust PGO: The system finds a maximum clique of mutually consistent measurements, passes the inliers to distributed PGO, and repeatedly exchanges involved pose estimates until consensus.

IV. EXPERIMENTAL RESULTS

The experiments span simulation, the KITTI00 benchmark, outdoor drone flights, and underground field tests.

  • The evaluation covers four settings: simulated multi-robot SLAM, KITTI00, outdoor football-field flights, and underground environments.

A. Implementation Details

DOOR-SLAM combines established robotics software with distributed communication, visual odometry, place recognition, feature processing, and robust pose graph optimization.

  • The system uses ROS for sensor interfacing and information exchange, and Buzz for multi-robot communication and scheduling.
  • Its front-end uses RTAB-Map for stereo visual odometry and TensorFlow NetVLAD descriptors, retaining only the first 128 dimensions to limit exchanged data.
  • Visual processing uses Good Features to Track with ORB descriptors, while distributed robust pose graph optimization is implemented in C++ with GTSAM.

B. Simulation Experiments

Simulation experiments evaluate outlier rejection and trajectory accuracy as PCM likelihood thresholds vary, using repeated runs in ARGoS.

  • Fig. 5 reports rejected inliers and outliers over 100-run averages with standard deviations in ARGoS.
  • Fig. 6 reports average translation error over 10-run averages with standard deviations for different PCM likelihood thresholds in ARGoS.
  • The simulation evaluates outlier rejection, inlier rejection, and average translation error as PCM thresholds change.Outlier rejection targets spurious measurements, inlier rejection measures needless removal, and ATE measures estimate distortion.
  • A lower PCM threshold rejects more measurements, including inliers, whereas a higher threshold can accept occasional outliers and increase error.

C. Dataset Experiments

Dataset and field experiments test DOOR-SLAM with aggressive loop-closure parameters, including KITTI00 evaluation and resource-constrained drone deployments.

  • C. Dataset Experiments: On KITTI00, three sequence parts run across three NVIDIA Jetson TX2s using a 1% PCM threshold, 0.15 NetVLAD threshold, and five feature correspondences.PCM enables more aggressive NetVLAD and geometric-verification settings than conservative outlier-avoidance parameters.
  • C. Dataset Experiments: 86.85m ATE without outlier rejection fell to 8.00m with PCM on KITTI00.The evaluation omitted intra-robot loop closures, making its error higher than recent SLAM solutions on this sequence.
  • C. Dataset Experiments: Field experiments used two stereo-camera quadcopters flying over a repetitive football field that induces perceptual aliasing.
  • C. Dataset Experiments: Recorded images and GPS data supported offline experiments on two WiFi-connected Jetson TX2s, followed by online onboard execution during flight.
  • C. Dataset Experiments: Fig. 9 compares accepted and rejected inter-robot loop closures across NetVLAD thresholds while fixing geometric verification at five feature correspondences.

1) Influence of Parameters:

DOOR-SLAM uses outlier-resilient back-end processing to permit less conservative front-end parameters, increasing valid loop closures while preserving trajectory quality.

  • Influence of Parameters: DOOR-SLAM’s back-end handles spurious measurements, making performance less sensitive to front-end parameter tuning.This robustness permits more loop closure candidates to be considered before outlier rejection.
  • Influence of Parameters: About three times more valid measurements are obtained with a NetVLAD threshold of 0.15 than with a threshold of 0.10.The looser threshold produces more candidates, although PCM rejects many of the additional loop closures.
  • Influence of Parameters: Reducing the minimum feature correspondences for successful geometric verification to 4 or 5 doubles valid inter-robot loop closures.RTAB-Map’s default is 20 correspondences.
  • Influence of Parameters: Higher PCM thresholds retain noisier loop closures, while thresholds above 75% or no PCM cause large increases in average translation error.A 1% PCM threshold yields the lowest average translation error in the analyzed environment.
  • Influence of Parameters: Using less conservative front-end parameters with a more conservative PCM threshold yields more valid loop closures and more accurate trajectory estimates.The front end gathers candidates, while PCM retains the most accurate measurements before optimization.

2) Communication:

The distributed front-end exchanges keypoint and relative-pose information rather than raw sensor data, reducing communication requirements while supporting online trajectory estimation.

  • Communication: The front-end shares keyframe keypoint information and sends relative pose transformations when enough feature correspondences are available.This supports distributed loop-closure detection without transmitting raw camera data.
  • Communication: DOOR-SLAM’s online trajectory estimates are reported alongside GPS ground truth used only for benchmarking.The figure shows estimates from both robots and the benchmark reference.
  • Communication: The proposed front-end reduces required bandwidth by roughly a factor of 10 compared with directly transmitting camera images.The comparison uses average message sizes measured during field experiments.

3) Online Experiments:

DOOR-SLAM was tested online with flying robots and in a lidar-based subterranean setup, where PCM improved mapping despite a remaining incorrect loop closure.

  • Online Experiments: Online experiments used two quadcopters running DOOR-SLAM on NVIDIA Jetson TX2 platforms with a 6Hz onboard camera.Visual odometry parameters were tuned for computational performance at some cost in accuracy.
  • Online Experiments: The online trajectory estimates were consistent with prior results despite degraded visual odometry accuracy.The experiment used a 1% PCM threshold, a 0.13 NetVLAD threshold, and five geometric-verification inliers.
  • Online Experiments: The lidar-based back-end was used with minimal modifications on data from two Husky UGVs in a coal mine over 1 kilometer long.Loop closures were detected by ICP scan matching in a self-similar environment prone to perceptual aliasing.
  • Online Experiments: PCM largely improved subterranean mapping, but one incorrect loop closure remained.The distributed back-end transmitted 92.27kB versus 196.30kB for the centralized setup, roughly halving communication burden.
  • Online Experiments: Across simulations, datasets, and field tests, DOOR-SLAM rejected spurious measurements and computed accurate trajectory estimates.The conclusion also reports successful use of less conservative front-end parameters.
Loading 1909.12198v2…