Multi-Robot Anomaly Detection

Period 2025.03 – 2026.05 · Affiliation KETI (multi-robot cooperative navigation national R&D project) · Role Fault injection and logging system development, anomaly detection framework design, model training and evaluation, web visualization

Background and Problem

In environments dense with multiple robots, a small fault in an individual robot (sudden battery drop, localization failure, navigation control failure) can propagate into fleet-wide bottlenecks and collisions. However, existing RMF focuses on visualizing individual robot states and lacked the means to diagnose inter-robot dependencies or cooperation failures, and threshold-based detection struggled to capture the irregularity of real anomaly data.

Approach

Phase 1 — Fault Injection and Data Collection

Built a PySide6 Fault Injection GUI that injects and recovers hardware (battery, wheel), Localization Lost, obstacle, and sensor (LiDAR, camera) faults on simulated robots.
Built a system that logs multi-robot information into a unified ROS 2 CSV (30 robots, 10 Hz downsampled to 1 Hz), later used as training data.

Phase 2 — Rule-Based Anomaly Detection and Web Visualization (ICCAS 2025) (Kim et al., 2025)

Anomaly Detection Framework — RMF (Task Allocation) → Anomaly State (4 types) → Web Dashboard → Operator.

Defined four anomaly types at the individual robot level (Battery Depletion, Obstacle Detection) and the fleet level (Trajectory Conflict, Collision Risk — paths crossing within 0.2 m of another robot).
In a modular structure where RMF handles task allocation and a separate sensor/log-based module handles anomaly detection, a WebSocket/ROS Bridge-based web dashboard delivered real-time alerts and visualization to the operator. Verified in a 30-robot Gazebo environment.

Phase 3 — Learning-Based Anomaly Detection (ICROS 2026) (Kim et al., 2026)

Open-RMF-based 10-robot simulation environment (Gazebo + RMF visualization).

Normalized each robot’s time series into a 4-dimensional feature \(X_t = [P_t, V_t, B_t, R_t]\) (position error, velocity, battery, mission progress).
After LSTM time-series prediction and Transformer AutoEncoder reconstruction-based detection, adopted RESTAD, a self-supervised model that embeds an RBF layer in a Transformer, to learn spatio-temporal correlations among multiple robots without labels.
Applied it to real robots through a real-time ROS 2 streaming pipeline (based on Motion Residual).

Results

Phase	Outcome
Rule-based (ICCAS 2025)	Real-time detection and visualization of four anomaly types, verified in a 30-robot simulation, first author, IEEE Xplore
Learning-based (ICROS 2026)	Region-level Aff-F1 0.94 — outperforming LSTM and a plain Transformer, first author
Operational visibility	PySide6 integrated fleet management GUI + Fault Injection GUI + web dashboard

Tech Stack

Open-RMF · ROS 2 · Gazebo · PySide6 / rclpy · WebSocket / ROS Bridge · PyTorch (LSTM, Transformer AE, RESTAD) · CSV logging

In an Open-RMF simulation with ten mobile robots, position error, velocity, battery level and mission progress are defined as time-series features, and a Transformer-based self-supervised model (RESTAD) learns spatio-temporal correlations across robots to detect anomalies without labels. On the region-level metric (Aff-F1 0.94) it outperforms LSTM and a plain Transformer. Learning-based follow-up to the rule-based ICCAS 2025 framework.

Defines four anomaly types in a multi-robot system (battery depletion, unexpected obstacle, trajectory conflict, collision risk within 0.2 m), detects them in a module decoupled from RMF task allocation, and visualizes them in real time on a WebSocket/ROS-bridge web dashboard. Validated in a Gazebo simulation with 30 mobile robots.