Edge-Assisted Multi-Robot Visual-Inertial SLAM with Efficient Communication

AI summary

Overview

Research area: Multi-robot Simultaneous Localization and Mapping (SLAM), specifically monocular Visual-Inertial SLAM (VINS) running across a robot–edge–cloud layered architecture.

Technical level: Advanced. The paper assumes familiarity with visual-inertial odometry, IMU pre-integration on manifolds, nonlinear graph optimization, ORB features, and lossless feature compression.

Scope: The paper proposes a centralized multi-robot visual-inertial SLAM system that splits feature extraction, VIO, and global optimization across a robot, an edge server, and a cloud server, and that compresses what is transmitted so that collaborative localization works under limited bandwidth.

What This Paper Is About

Multi-robot SLAM needs robots to share data so they can localize jointly and build one consistent global map, but the bandwidth between robots and a central server is limited and communication links are long, which degrades performance. The authors' goal is to keep collaborative SLAM globally consistent and real-time while transmitting as little data as possible, without losing accuracy to compression. They attack this from two sides: a lightweight front end that avoids computing descriptors on most frames, and a layered robot–edge–cloud architecture that decouples where each computation happens.

Key Contributions

1. A lightweight SLAM method built on ORB-SLAM3 and Lucas–Kanade (LK) sparse optical flow. Descriptors are computed only at keyframes; pose estimation between non-keyframes uses sparse optical flow tracking guided by IMU prediction inside an image pyramid. This reduces the cost of feature tracking without affecting the accuracy of back-end optimization.

2. A centralized multi-robot SLAM system on a robot–edge–cloud hierarchical architecture. The classic SLAM framework is decoupled, with feature extraction, Visual-Inertial Odometry (VIO), and global optimization placed in separate layers, avoiding the limited onboard computing resources and low execution efficiency of a single robot.

3. Lossless coding and compression of corner points (non-keyframes) and local binary descriptors (keyframes) within the robot–edge–cloud architecture, enabling real-time remote transmission under limited bandwidth. The design reduces occupied bandwidth without causing the loss of SLAM accuracy that data compression can introduce.

4. Introduction of map backbone profiling (MBP) into the method to guarantee optimization efficiency and reduce storage.

Main Findings

• Reduced transmitted data volume: On the EuRoC dataset, compared with the current most advanced local feature compression method, the proposed method achieves lower data volume feature transmission.
• Comparable or better accuracy: Compared with the current advanced centralized multi-robot SLAM scheme, it achieves the same or better positioning accuracy under low computational load.
• No compression-induced accuracy loss: Because encoding is lossless by design, the compression itself does not degrade SLAM accuracy.
• Descriptor cost is confined to keyframes: ORB feature extraction takes nearly 20 ms, which the authors describe as a challenge to onboard computing power; in this system descriptor calculation and matching occur only at keyframes, while LK sparse optical flow handles inter-frame tracking for non-keyframes using constant-luminosity, small-motion, and locally-consistent-motion assumptions.
• IMU replaces camera-based motion models: Where FastORB-SLAM uses a uniform acceleration motion model from camera information, this work uses IMU pre-integration for pose prediction, which the authors argue is more robust in dynamic scenes. Before IMU initialization completes, a uniform motion model is used, with velocity computed from mean angular velocity in the IMU frame.
• Pyramid and coding parameters: The pyramid uses scale ratio s = 1.2 and n = 8 layers. Keypoint coding uses N_sigma = 8 pyramid levels and N_theta = 32 angular bins for keyframe keypoints.
• Edge–cloud traffic composition: Keyframe data — compressed 2D feature points — accounts for the main traffic between edge and cloud, with each compressed 2D feature key occupying about 220 bytes. Smaller items include IMU measurements, current bias and velocity estimates, and confirmation messages for keyframe and map point updates.
• Reliable transport: Communication uses socket programming over TCP with a message packet-loss retransmission mechanism, and clients create connection sockets, so robots, edge servers, and the cloud can join and request communication dynamically during a task.
• Quantitative benchmark numbers are not reported in the provided content. The truncated text names the EuRoC dataset but does not give trajectory error figures, bandwidth savings percentages, or dataset sizes.

Methodology in Plain English

The system is organized into three layers.

On the robot, the system detects features quickly. For non-keyframes it uses LK sparse optical flow with IMU-assisted prediction rather than computing descriptors; for keyframes it extracts ORB features. Non-keyframes and keyframes are encoded differently: a non-keyframe keypoint is stored as horizontal and vertical coordinates plus pyramid level, while a keyframe keypoint adds an orientation angle. Keyframes also carry a visual word index from the ORB-SLAM3 vocabulary (shared between robot and edge) and a binary descriptor. For the descriptor, the encoder finds the closest visual word and encodes only the XOR residual vector, whose bit cost depends on the descriptor length D, the number h of non-zero residual elements, and the probability p0 that a residual element is zero.

On the edge server, the VIO front end runs. It solves the current frame's pose from optical flow tracking for non-keyframes, and falls back to the ORB-SLAM3 tracking method for keyframes when optical flow tracking fails. The edge maintains a limited-size local map for each agent so each agent retains basic autonomy. IMU initialization runs in the VIO local-mapping thread so inertial parameters are more precise, and after initialization the edge sends the better inertial parameters back to the robot.

On the cloud, the expensive, global, non-real-time work happens: global pose graph optimization and global bundle adjustment, redundant keyframe culling, loop closure, and global map fusion. The edge notifies the cloud of keyframe and map changes; the cloud updates keyframe poses and the global map and sends updated keyframe poses back to the edge, which quantifies drift in the current attitude estimate.

Initialization proceeds through pure-visual optimization, then only-inertial optimization, then visual-inertial joint optimization, with scale and gravity optimized while keyframe poses and parameters stay fixed.

Why This Matters

Impact on research. The paper targets a recognized bottleneck in multi-robot SLAM: bandwidth and communication stability, which the authors note remain under-studied for centralized systems operating over unstable links. It also extends an edge-assisted architecture that was previously introduced for 2D LiDAR robots into the visual-inertial setting, and it separates what must be transmitted for keyframes versus non-keyframes rather than treating all frames alike.

Real-world applications (drawn from the scenarios the paper cites for SLAM and IoT):

• Mobile robots operating in complex or large-scale environments.
• Vehicles, where the authors' affiliated work connects to intelligent transport research.
• Underwater robots.
• Internet of Things deployments that need low delay, high bandwidth, high reliability, massive connections, heterogeneous aggregation, and local security and privacy protection at the edge.

Industry relevance. The approach lets cheaper, lighter robots do less computation while a shared edge and cloud do the heavy optimization, which suits fleets with constrained onboard computing, cost, and payload. Because the encoding is lossless and descriptor computation is confined to keyframes, the bandwidth savings do not come at the price of map accuracy — a trade-off that matters for commercial deployments on metered or unreliable networks.

Future Directions

• Apply the framework to more real-world systems rather than only public datasets, as the authors state they intend to do.
• Achieve rich, accurate map fusion using more advanced features, which the authors identify as a goal for future study.
• Study centralized SLAM systems under unstable communication, which the paper explicitly identifies as still to be researched.
• Extend the architecture further: the authors note that existing edge-assisted SLAM work mostly covers single-robot, single-edge settings and cannot be directly extended to communication between multiple robots and multiple edges.

Target Audience

Researchers and engineers working on multi-robot SLAM, visual-inertial odometry, and cloud/edge robotics; practitioners building fleets of robots over constrained or unreliable networks; and edge-computing researchers interested in how to split perception and optimization workloads across robot, edge, and cloud tiers. Readers need a background in SLAM estimation and feature-based vision to follow the pre-integration and encoding derivations.

Authors’ abstract

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.