Abstract
Traditional decentralized cooperative localization (DCL) methods often rely on precise sensor noise models and rigid statistical assumptions, which fail in the presence of sensor faults or high-bias environments. This paper introduces Distributed Learning-based Decentralized Cooperative Localization (DL-DCL), an algorithm that utilizes a moving landmark as a feedback source for real-time online learning. DL-DCL achieves a 40% improvement in localization accuracy compared to standard industry methods like Covariance Intersection (CI) without requiring prior covariance knowledge.
Introduction
In autonomous robotics, localization is critical for navigation and task execution. While individual robots use Inertial Measurement Units (IMUs) and GPS, these sensors are prone to failure or significant bias in challenging environments. Cooperative Localization (CL) allows robots to share data to improve collective positioning. However, conventional DCL algorithms, such as Kalman Filters, are sensitive to "noise" and require accurate mathematical descriptions of uncertainty (covariance). When a sensor experiences a sudden 50x increase in bias, these models typically fail. DL-DCL addresses this by providing a fault-tolerant mechanism for robots to learn trust weights for sensor data in real-time.
Methods
The DL-DCL framework operates in a decentralized and distributed manner, where computation is spread across the robot fleet. The system architecture involves multiple robots equipped with IMUs and relative pose sensing (RPS) capabilities, interacting with a moving landmark that broadcasts its real-time position.
The algorithm consists of three primary phases:
- Projection Phase: Robots predict their future state based on previous estimates and movement commands.
- Learning Phase: Robots observe the moving landmark and compare their estimated landmark position with the actual broadcasted position. This discrepancy generates a "loss" value, used to update weights via an exponentially weighted online learning rule.
- Estimation Phase: Robots fuse their local sensor data with neighbor data using the learned weights to produce a high-accuracy pose estimate.
Key Results
The performance of DL-DCL was evaluated against traditional methods including Kalman Fusion (KF), Covariance Intersection (CI), and Covariance Union (CU). In simulations where sensors were intentionally "broken" with 40-60x nominal bias, DL-DCL demonstrated superior robustness.
Key findings include:
- Accuracy: DL-DCL showed a ~40% reduction in estimation error compared to CI.
- Fault Tolerance: The learning mechanism effectively ignored "garbage" data from failed sensors.
- Efficiency: With a per-iteration time of ~1.2 ms, the algorithm is suitable for real-time applications.
- Scalability: System reliability improved as the number of robots increased, maintaining stable error rates.
Conclusion
DL-DCL provides a robust, model-free approach to cooperative localization. By replacing rigid covariance assumptions with dynamic learning from moving landmarks, the system achieves high accuracy and fault tolerance. Future work will focus on extending the model to 3D environments and addressing challenges related to landmark visibility and extreme communication failures.
Summary updated