Source-linked AI summary
Edge-Assisted Multi-Robot Visual-Inertial SLAM with Efficient Communication
Xin Liu, Shuhuan Wen, Jing Zhao, Tony Z. Qiu, Hong Zhang
TL;DR
Limited bandwidth and long terminal-to-cloud links can degrade real-time multi-robot SLAM. The paper proposes lightweight IMU-assisted optical-flow tracking and centralized robot-edge-cloud processing with lossless feature compression. Experiments report lower transmitted feature data and positioning accuracy consistent with, or better than, advanced centralized methods under low computational load.
Problem
Limited bandwidth and long communication links can degrade multi-robot SLAM performance, while non-keyframe descriptor computation consumes substantial resources.
Method
The paper combines IMU-assisted LK optical-flow tracking between non-keyframes with centralized robot-edge-cloud SLAM and lossless feature coding.
Results
EuRoC experiments show lower transmitted feature data and positioning accuracy consistent with, or better than, advanced centralized methods under low computational load.
Takeaways & Limitations
The system supports global sparse-map construction and joint positioning through multiple independent robots with low transmitted data.
Abstract
from arXiv · showhide
The integration of cloud computing and edge computing is an effective way to achieve global consistent and real-time multi-robot Simultaneous Localization and Mapping (SLAM). Cloud computing effectively solves the problem of limited computing, communication and storage capacity of terminal equipment. However, limited bandwidth and extremely long communication links between terminal devices and the cloud result in serious performance degradation of multi-robot SLAM systems. To reduce the computational cost of feature tracking and improve the real-time performance of the robot, a lightweight SLAM method of optical flow tracking based on pyramid IMU prediction is proposed. On this basis, a centralized multi-robot SLAM system based on a robot-edge-cloud layered architecture is proposed to realize real-time collaborative SLAM. It avoids the problems of limited on-board computing resources and low execution efficiency of single robot. In this framework, only the feature points and keyframe descriptors are transmitted and lossless encoding and compression are carried out to realize real-time remote information transmission with limited bandwidth resources. This design reduces the actual bandwidth occupied in the process of data transmission, and does not cause the loss of SLAM accuracy caused by data compression. Through experimental verification on the EuRoC dataset, compared with the current most advanced local feature compression method, our method can achieve lower data volume feature transmission, and compared with the current advanced centralized multi-robot SLAM scheme, it can achieve the same or better positioning accuracy under low computational load.
I. INTRODUCTION
The paper motivates multi-robot SLAM for complex scenes and proposes a lightweight visual-inertial system that addresses communication, computation, and real-time constraints through collaborative architecture and compressed feature processing.
- Motivation: Multi-robot collaboration improves task execution efficiency, robustness, flexibility, and fault tolerance over single-robot SLAM in complex and large scenes.
- Motivation: Centralized architectures preserve global data consistency and balance computing load, but communication efficiency and data volume can bottleneck real-time performance.
- Contributions: The paper proposes a fast, lightweight, real-time multi-robot monocular Visual-Inertial SLAM system.
- Contributions: The lightweight method computes descriptors only on keyframes and uses IMU-predicted LK sparse optical flow for pose estimation between non-keyframes.
- Contributions: A robot-edge-cloud hierarchy separates feature extraction, VIO, and global optimization to address limited onboard resources and single-robot execution inefficiency.
- Contributions: Lossless encoding and compression transmit non-keyframe corner points and keyframe local binary descriptors under limited bandwidth without compression-induced SLAM accuracy loss.
B. Muti-Robot SLAM
Prior multi-robot SLAM systems trade scalability, communication load, consistency, and computational cost. The paper positions edge-assisted centralized processing and analysis-before-compression as a response to these constraints.
- Multi-robot architectures: Decentralized systems scale readily and compute in parallel, but asynchronous data and increased network load challenge mapping accuracy and localization.
- Multi-robot architectures: Centralized systems support global data consistency and task-based computing-load balancing, making them common in collaborative visual SLAM.
- Edge-assisted SLAM: Edge computing moves workload from the cloud to physically nearby servers, but earlier edge-assisted SLAM systems focused on single-robot settings.
- Edge-assisted SLAM: Analysis followed by compression extracts and transmits local features instead of complete images, reducing bandwidth relative to full-image transmission.
- Proposed architecture: The proposed robot-edge-cloud architecture extends edge-assisted processing to Visual-Inertial multi-robot SLAM by analyzing and then compressing sensor data.
IV. IMU-ASSISTED NON-KEYFRAMES FEATURE MATCHING
The method uses IMU prediction to guide sparse LK optical-flow matching between non-keyframes, avoiding descriptor computation while retaining pose-estimation inputs for visual-inertial tracking.
- Motivation: Descriptor computation and matching are costly, so the method performs LK sparse optical-flow tracking between non-keyframes and computes descriptors only in keyframes.This preserves keyframe descriptors for loop detection and multi-robot map fusion while reducing inter-frame tracking cost.
- Optical-flow assumptions: Sparse optical flow assumes constant luminosity, small inter-frame object motion, and locally consistent pixel motion.The constant-luminosity relation matches corresponding points by their gray values across frames.
- Pose prediction: Before IMU initialization, the uniform motion model estimates pose using mean angular velocity from consecutive IMU measurements.Translation is neglected over the short interval; after initialization, pose estimation switches to IMU pre-integration.
- IMU model: IMU measurements are modeled with gyroscope and accelerometer noise and biases, whose random-walk errors follow a Wiener process.The model distinguishes body and world frames and includes gravitational acceleration in the accelerometer relation.
- IMU pre-integration: IMU pre-integration between frames produces rotation, velocity, and position increments used to initialize visual estimation.Integrating all IMU information between consecutive frames provides the next frame’s PVQ initial value.
- Tracking pipeline: The tracking pipeline predicts current-frame keypoint pixels with the estimated transform, runs pyramid LK tracking, filters matches, and solves pose with PNP.The algorithm tracks from the top pyramid level downward, screens points using IMU rotation, removes outliers with RANSAC, and returns Tcw.
V. METHOD OF COMMUNICATION
The communication method uses an edge-assisted layered architecture to support accurate positioning and sparse-map establishment without transmitting non-keyframe descriptors.
- Communication architecture: The cloud-edge layered architecture enables accurate positioning and sparse-map establishment without transferring non-keyframe descriptors.This minimizes the information transferred during multi-robot SLAM.
A. Feature Coding Compression
The feature-coding scheme uses lossless intra-frame encoding with separate keyframe and non-keyframe modes, transmitting only keypoints for non-keyframes and richer feature data for keyframes.
- Encoding modes: The lossless intra-frame encoder divides transmission into keyframe and non-keyframe modes.Keyframe selection depends on tracked-point count and elapsed time from the previous keyframe.
- Encoding modes: Non-keyframe encoding transmits only keypoints, while keyframe encoding includes ORB visual words, keypoints, and descriptors.The scheme therefore avoids transmitting non-keyframe descriptors.
- Encoding cost: The encoding-cost expressions separately account for non-keyframe keypoints and keyframe keypoints, visual-word indices, and descriptors.The notation distinguishes Ckf,BoW and Ckf,des from the non-keyframe and keyframe keypoint costs.
- Keypoint coding: A non-keyframe keypoint is represented by its horizontal coordinate, vertical coordinate, and pyramid hierarchy.These three properties define the compact non-keyframe keypoint information.
- Keypoint coding: Keyframe keypoints additionally encode orientation, with pyramid-frame coordinates replacing image-frame coordinates to reduce bandwidth.The orientation is quantized into Nθ parts, while pyramid layers use Nσ = 8 and Nθ = 32.
2) Compression encoding of visual word index and descriptor:
Visual-word and descriptor coding combines a shared ORB vocabulary with lossless residual encoding based on descriptor XOR differences.
- Visual-word coding: Each feature is assigned the visual-word index with the smallest Hamming distance in the shared ORB visual-word tree.The visual-word index is then encoded for transmission.
- Descriptor coding: The descriptor residual is computed by XOR between the ORB descriptor and its closest visual word for lossless representation.The residual encoding cost depends on descriptor length, the number of non-zero elements, and the zero-element probability p0.
B. Robot-Edge-Cloud Communication
The system uses reliable robot–edge–cloud communication and lightweight visual-inertial processing to support collaborative SLAM under constrained resources.
- Communication design: TCP-based socket communication uses retransmission to avoid packet loss and preserve message reliability.
- Communication design: Keyframes and non-keyframes are encoded and compressed separately to reduce robot–edge bandwidth usage.
- Visual-inertial processing: IMU initialization runs on the local mapping thread, after which the edge sends refined inertial parameters to the robot.
- Edge–cloud communication: The edge continuously sends updated keyframes and mappoints to the cloud, with each compressed 2D feature key occupying about 220 bytes.
- Visual-inertial processing: Visual-inertial joint optimization refines scale and gravity after separate pure-visual and inertial initialization estimates.
- Local optimization: Local bundle adjustment combines camera reprojection and IMU pre-integration residuals with robust Cauchy and Huber losses.
B. Map Fusion
After edge-side local mapping, the cloud performs global pose-graph optimization across robots and mappoints while preserving robot-specific inertial constraints.
- Global optimization: Cloud-side GPGO detects candidate keyframes with DBoW and refines relative poses through feature matching and reprojection-error optimization.
- Loop closure: Global optimization treats intra-robot and inter-robot loop closures uniformly within the same cloud-based process.
- Global optimization: GPGO optimizes all keyframes and mappoints, whereas IMU constraints connect only successive keyframes belonging to the same robot.
- Scalability: Scene growth increases GPGO computation, motivating storage reduction while avoiding pruning choices that can remove well-observed mappoints or reduce final accuracy.
VII. EXPERIMENTS
The system is evaluated on EuRoC sequences spanning a machine hall and two indoor rooms, using calibrated camera–IMU data and laser-tracker ground truth on real devices.
- Sensors and ground truth: Camera data come from an Aptina MT9V034 global-shutter camera, while IMU data come from an ADIS16448.
- Sensors and ground truth: Ground truth is captured with a Leica MS50 laser tracker, and camera and camera–IMU parameters are obtained using Kalibr.
- Hardware platform: Experiments run on real devices with an Intel i3-9100 robot, an Intel i7-9750H edge server, and an AWS cloud server.
A. Optical Flow Tracking Accuracy and Speed
The proposed IMU-assisted optical-flow front end improves feature tracking in most EuRoC frames while reducing transmitted feature data relative to a local compression method.
- Tracking method: The front end combines ORB-SLAM3 with LK optical flow and avoids descriptor computation for non-keyframe matching.
- Tracking accuracy: IMU-assisted tracking generally matches more points and produces fewer mismatches than traditional LK optical flow on EuRoC.
- Tracking speed: In most consecutive EuRoC frames, the proposed method improves both the number of tracked points and tracking speed over LK optical flow.
- Tracking caveats: The tracking gain is limited under significant illumination changes and is not stable across every frame because of RANSAC’s random selection.
- Communication rate: 46.81kbits/s is the proposed robot-to-edge communication rate across eight EuRoC sequences, versus 211.9kbits/s for method.
C. Localization and Map Fusion Accuracy
The proposed centralized system evaluates collaborative positioning using ATE RMSE on EuRoC sequences and compares its accuracy with existing methods. Results also examine feature compression and redundant keyframe removal, while V2 trajectories are aligned with reference groundtruths.
- Accuracy evaluation: ATE RMSE evaluates collaborative positioning accuracy and global trajectory consistency across MH 01-MH 03, MH 01-MH 05, V1 01-V1 03, and V2 01-V2 03.ATE directly measures differences between estimated and groundtruth poses.
- Accuracy evaluation: The proposed method achieves average RMSE across all sequences that is consistent with the leading centralized method and better than most reported results.The comparison uses results from CVIDS, COVINS AWS, VINS-mono, and ORB-SLAM3.
- Accuracy evaluation: MBP improves performance over the proposed system without MBP in all four experiment groups, indicating gains for edge-assisted multi-robot SLAM.The paper separately reports poor localization on MH 01-MH 05 with COVINS, while noting faster optical-flow front-end processing.
- V2 evaluation: In the challenging V2 scenario, each robot trajectory and the combined trajectory align well with the reference groundtruths.The visualization distinguishes tracks and map points across three sequences and marks common views or closed-loop keyframes.
D. Real scene experiment
The real-scene experiment deploys the system on heterogeneous TurtleBot platforms with depth cameras and independently measured groundtruth. The estimated trajectories form a coherent map and show good agreement with the real-scene reference.
- Experimental setup: The experiment uses two TurtleBot 3 robots with Raspberry Pi 3 computers and one TurtleBot 2 with a Lenovo Y7000 computer.All robots use MYNT EYE depth cameras, with camera-IMU calibration performed by kalibr.
- Experimental setup: Marvelmind indoor GPS and IMU measurements provide the reference groundtruth, using fixed anchors to locate moving anchors.The groundtruth is calculated by fusing the positioning and inertial information.
- Real-scene results: The three robots follow different trajectories, meet at the endpoint, and form an obvious covisible relationship between the start and end points.The robots circle at the starting point for two weeks before completing the closed loop.
- Real-scene results: The proposed method produces trajectories that align with the reference groundtruth and shows good estimation performance in the actual scene.The estimated trajectory is also overlaid on the real scene’s 2D grid map for visual comparison.