Decentralized Cooperative Localization for Multi-Robot Systems with Asynchronous Sensor Fusion

Multi-robot systems have demonstrated widespread applications across diverse fields including surveillance, search and rescue, space exploration, agriculture, logistics, and military operations. These systems efficiently cover extensive areas, perform coordinated tasks, and provide enhanced robustness compared to single-robot approaches. Cooperative localization is a key issue in multi-robot systems, especially in GPS-denied environments. Traditional centralized methods, although theoretically optimal, face challenges such as communication bandwidth constraints, single-point-of-failure risks, and scalability issues. Decentralized methods distribute computation across robots, offering enhanced scalability and robustness.

In GPS-denied environments, traditional centralized cooperative localization methods face challenges such as communication bandwidth constraints, single-point-of-failure risks, and scalability issues. Additionally, these methods often assume known globally aligned coordinate frames, synchronous or near-synchronous sensor sampling, and continuous communication bandwidth for state exchange. However, these conditions are rarely satisfied in real-world scenarios like underground mines, subsea operations, or enclosed industrial facilities.

This paper proposes a decentralized cooperative localization (DCL) framework with the following innovations:

1. Automatic alignment through transformation matrices: Allows robots to initialize with arbitrary reference-frame orientations, eliminating the need for global alignment or strict synchronization.

2. Dual-landmark evaluation framework: Leverages both static environmental features and mobile robots as dynamic landmarks to improve observability in feature-sparse scenes.

3. Asynchronous sensor fusion: Utilizes ROS message filters for timestamp-based fusion of asynchronous sensor streams, ensuring temporal consistency.

4. Event-triggered information exchange: Shares information only during mutual observations, reducing bandwidth usage by 65%.

The methodology of this paper includes the following steps:

• �� Use an Extended Kalman Filter (EKF) for local localization, with each robot maintaining its pose estimate, self-covariance, and cross-covariance.
• �� In the prediction stage, receive encoder data and update state estimates and covariances.
• �� In the measurement and update stage, use transformation matrices for measurement model conversion, achieving alignment with arbitrary reference frames.
• �� Employ the Adaptive Breakpoint Detector (ABD) for reliable feature extraction, ensuring robustness under sensor noise and environmental clutter.
• �� Implement asynchronous data synchronization using timestamp buffering to ensure temporal consistency of sensor data.
• �� Integrate a dual-landmark strategy, utilizing both static and dynamic landmarks for continuous corrections.

The experimental design includes validation in both Gazebo simulation and real-world environments. Two differential-drive mobile robots operate in a 2D space for 100 seconds with a timestep of 0.1 seconds. The methods compared include centralized cooperative localization (CCL), decentralized cooperative localization (DCL), and the dual-landmark variant (DCL-LM). The experimental environment simulates GPS-denied conditions with weak WiFi signals, concrete walls, and limited visual landmarks.

Experimental results show that DCL outperforms centralized cooperative localization (CCL) in both Gazebo simulation and real-world environments, achieving a 34% reduction in RMSE, while the dual-landmark variant yields a 56% improvement. In feature-sparse environments, the dual-landmark strategy significantly improves detection and feature extraction reliability by utilizing both static and dynamic landmarks. During dynamic maneuvers, the DCL framework effectively handles asynchronous sensor data while maintaining cross-correlation consistency among robot state estimates.

The DCL framework is applicable to challenging domains like underground, underwater, and feature-sparse terrains. In these environments, traditional localization methods often fail, while the DCL framework enhances system robustness and scalability through a decentralized approach. Specific application scenarios include underground mine inspection, underwater operations, warehouse automation, and disaster response in enclosed structures.

Despite the advantages of the DCL framework, localization accuracy may degrade during prolonged communication interruptions. Additionally, geometric ambiguity may lead to feature extraction failures when robots are very close to walls (less than 0.3 meters). Future research directions include scaling to larger teams through structured cross-covariance management, integrating complementary asynchronous sensors (IMU, visual odometry, UWB) with adaptive noise tuning.