Open-RMF Multi-Robot Fleet Management

Period 2023.05 – present · Affiliation Intelligent Robotics Research Center, Korea Electronics Technology Institute (KETI) · Role Design and real-robot validation of the multi-robot simulation and fleet management system

Background and Problem

For dozens of mobile robots with different control systems to operate together in a large indoor workspace, a supervisory architecture is needed that assigns missions centrally and runs the fleet without collisions. When many robots run simultaneously, global path interference and server communication bottlenecks were the core challenges. This project covers the full process of unifying heterogeneous robots under one fleet manager based on Open-RMF, and transferring the logic verified in simulation to physical robots.

Responsibilities

Multi-robot simulation infrastructure: Extended Gazebo models from TurtleBot3 to TETRA_S/TETRA_M with namespace and TF design, Nav2 multi-robot configuration, ROS 1 to ROS 2 (Foxy/Humble) migration, and sensor driver cleanup (LiDAR, IMU, RealSense).
Open-RMF fleet management and task system: Built traffic_editor nav_graphs, integrated fleet_adapter and free_fleet, implemented user-input-driven go_to_place dispatch, a node that automatically assigns endlessly repeating missions (RMF Infinity Task), and a battery-plugin-based charging task.
Cooperative perception and navigation: Each robot converts obstacles detected with a 2D LiDAR and RGB-D camera (YOLOv8) into global map coordinates and shares them with the server; lanes intersecting an obstacle are closed (Lane Closed) so that other robots proactively detour around obstacles outside their own field of view.
Communication optimization: Resolved communication bottlenecks and data loss in real-robot operation through CycloneDDS configuration, and extended connections to robots on other domains with a Zenoh bridge. Demonstrated image_transport-based fleet management in a WiFi-6E environment.
Path planning analysis: Analyzed the structure of centralized (CBS) and decentralized (VO, RVO, ORCA) MAPF algorithms and verified them in Gazebo.
Handover and integration testing: State-based Replan Trigger ROS 2 service, ROS 2 emergency-stop topic, server-centric TF restructuring, yearly integration tests, and definition of performance metrics and data collection for certified testing.

Results

Verified the fleet management logic in a 30-robot multi-robot simulation (RMF Gazebo World conversion, LiDAR avoidance, YOLO obstacle sharing), then completed a Sim-to-Real demonstration in which 4 physical robots executed missions without deadlock.
Confirmed through 4-robot experiments that cooperative obstacle information sharing reduces collision risk and path cost compared with a non-cooperative system (ICCAS 2024).
Related papers: KROC 2024 (Kim et al., 2024), ICROS 2024 (Kim et al., 2024), ICCAS 2024 (Kim et al., 2024) — first author on all three.

Full Control Fleet Adapter ↔ ROS 2 Free Fleet Server ↔ Free Fleet Client (ROS 1/ROS 2 mobile robots) architecture (KROC 2024).

Left: 2D LiDAR sampling + RGB-D YOLOv8 obstacle detection (ICROS 2024). Right: Robot (State/Navigation/Object Detection) ↔ RMF (Task/Traffic/Schedule) framework (ICCAS 2024).

Tech Stack

Open-RMF · ROS 2 (Foxy/Humble) · Nav2 · Cartographer · Gazebo · CycloneDDS / Zenoh · YOLOv8 · TETRA_S / TETRA_M · C++ / Python

A server-client fleet management system that integrates the Open-RMF middleware with ROS Navigation, so that multiple mobile robots receive tasks from the server and drive cooperatively without collisions using RMF traffic negotiation. Feasibility was validated with real-robot experiments.

Obstacles detected by each mobile robot with a 2D LiDAR and an RGB-D camera (YOLOv8) are transformed into the global map frame and shared so that other robots can use them for collaborative path planning. In the Open-RMF fleet manager, lanes are marked active or inactive from the shared obstacle information so that robots only travel on active lanes. Validated in Gazebo with two robots.

Each robot detects dynamic obstacles (people) with an RGB-D camera and YOLOv8 and shares them with the RMF server. When the number of obstacles adjacent to a graph edge exceeds a predefined threshold, the edge is deactivated and affected robots proactively re-plan. Experiments with four robots show reduced collision risk and path cost compared with a non-cooperative system.

Background and Problem

Responsibilities

Results

Tech Stack

References

2024