Improving Odometric Model Performance Based on LSTM Networks.

Area of Science:

• Robotics
• Artificial Intelligence
• Sensor Fusion

Background:

• Accurate localization is crucial for autonomous systems like wheelchairs.
• Traditional sensor fusion methods can be limited by static covariance assumptions.
• Odometry often suffers from drift and systematic errors.

Purpose of the Study:

• To develop an improved localization system for autonomous wheelchairs.
• To enhance odometric accuracy using deep learning techniques.
• To enable real-time self-correction and error detection in wheelchair navigation.

Main Methods:

• Utilized an LSTM neural network for odometric pose estimation.
• Trained the network using data from odometric encoders and Velodyne sensor ground truth.
• Implemented real-time network retraining for adaptive error correction.
• Developed a secondary network to detect non-systematic errors using motor power and wheel speed data.

Main Results:

• The proposed LSTM-based odometric model significantly improved localization accuracy.
• The system demonstrated real-time self-correction capabilities against temporal variations.
• Non-systematic errors were effectively detected using motor power and wheel speed correlations.
• The final robot localization outperformed classic sensor fusion with static covariance.

Conclusions:

• Deep learning, specifically LSTM networks, can substantially enhance odometric localization accuracy in autonomous wheelchairs.
• Real-time adaptation and error detection mechanisms improve the robustness and reliability of autonomous navigation systems.
• The developed system offers a more accurate and adaptive solution for robot localization compared to conventional approaches.