A Robust Multi-Sensor Fusion Localization for Small-Scale Biomimetic Robots

Robust localization is an important foundation for quadruped robots to achieve motion performance and application potential in real-world scenarios. Vision-based methods commonly suffer from undesired performance, especially in low-lighting conditions. To solve this problem, we present a multi-sensor fusion localization approach for slippery surfaces and poor visual quality conditions. The proposed approach creates an adaptive feature extraction framework that only extracts effective local features rather than global ones, improving data processing speed while ensuring sufficient feature extraction. Besides, we implement leg and inertial odometry based on motor encoders and IMU mounted on a small-scale robotic rat. The approach with information fusion has been proved with high robustness by reducing localization error by over 20% and processing time by 15.8%. The evaluation across multiple stages has been done and the results demonstrate competitive performance on both public benchmarks and a miniature quadruped robot, which shows high versatility for small-scale biomimetic robots in real-world scenarios.