Pose estimation of differential drive robots using deep learning and raw sensor inputs.

Area of Science:

• Robotics
• Machine Learning
• Sensor Fusion

Background:

• Accurate mobile robot localization is crucial for navigation and task execution.
• Existing methods often rely on feature extraction, which can be complex and time-consuming.
• Directly using raw sensor data offers a potentially more efficient and robust approach.

Purpose of the Study:

• To develop and evaluate an estimation method for mobile robot position and orientation using raw Inertial Measurement Unit (IMU) data and simulated velocities.
• To compare the performance of various machine learning models, including Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), Gradient Boosting (GB), and Random Forest (RF).
• To investigate the effectiveness of the proposed method on both simulated and real-world robot data.

Main Methods:

• Collected data from a TurtleBot3 mobile robot in ROS-Gazebo, including simulated and real-world routes.
• Incorporated real IMU sensor noise and pure pursuit algorithm velocities into the dataset.
• Trained and compared CNN, LSTM, GB, and RF models for position and orientation estimation.
• Evaluated models using raw sensor data without feature extraction.

Main Results:

• The Convolutional Neural Network (CNN) architecture consistently outperformed other models in estimating robot position and orientation across all tested routes.
• The proposed method demonstrated effectiveness in both simulated and real-world experimental scenarios.
• Direct utilization of raw sensor data proved viable and effective for robot localization.

Conclusions:

• The CNN-based approach offers a highly accurate and efficient method for mobile robot localization using raw IMU data and simulated velocities.
• This research contributes a novel technique that bypasses traditional feature engineering, simplifying the localization process.
• The findings suggest a promising direction for real-time robot localization systems.