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

## TL;DR

This paper introduces a novel method for estimating a robot's position and orientation using raw sensor data and deep learning, achieving better accuracy than other models.

## Contribution

The novelty lies in using raw sensor data directly in deep learning models without feature extraction for pose estimation.

## Key findings

- CNN models outperformed LSTM, GB, and RF in estimating robot position and orientation.
- Incorporating real IMU noise improved model accuracy in both simulated and real-world scenarios.
- The approach avoids feature extraction, simplifying the estimation pipeline.

## Abstract

This paper presents an estimation method for determining the position and orientation of a real mobile robot using raw data from an Inertial Measurement Unit (IMU) sensor, alongside linear and angular velocities obtained from simulation. The dataset was collected using a real TurtleBot3 differential drive wheeled mobile robot in the ROS-Gazebo simulation environment, encompassing 2018 routes-2009 from simulation and 9 from real-world experiments-each consisting of five randomly generated waypoints. To improve the accuracy of the estimation models, noise from the real IMU sensor was incorporated into the input data, and velocities derived from the pure pursuit algorithm were also included. Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), Gradient Boosting (GB), and Random Forest (RF) models were employed to estimate the robot’s position and orientation, and their performance was compared across both simulated and experimental scenarios. The results indicate that the CNN architecture consistently outperforms other models across all routes. Unlike many existing studies, this work directly utilizes raw sensor data without applying any feature extraction techniques, highlighting its novelty and contribution to the field.