NeoSLAM: Long-Term SLAM Using Computational Models of the Brain

Area of Science:

• Robotics
• Computer Vision
• Computational Neuroscience

Background:

• Simultaneous Localization and Mapping (SLAM) is crucial for robot navigation but camera-based methods struggle with long-term accuracy in diverse conditions.
• Existing visual SLAM systems face challenges in maintaining robust localization over extended periods and in complex environments.

Purpose of the Study:

• To introduce NeoSLAM, a novel long-term visual SLAM system inspired by neuroscience.
• To address the limitations of current visual SLAM methods in maintaining accurate localization over time and under challenging conditions.

Main Methods:

• NeoSLAM utilizes computational models of the human neocortex, specifically a hierarchical temporal memory model.
• It employs sparse distributed representations to identify temporal sequences of spatial patterns, offering high representational capacity and noise tolerance.
• A novel neuroscience-based loop-closure detector is developed for real-time performance, suitable for resource-constrained systems.

Main Results:

• The proposed NeoSLAM system demonstrated improved loop-closure detection accuracy compared to the traditional RatSLAM system.
• Evaluations were conducted using a wheeled robot in environments of varying complexity.
• The neuroscience-based approach proved effective in enhancing SLAM performance.

Conclusions:

• NeoSLAM offers a promising neuroscience-based solution for long-term visual SLAM.
• The system's reliance on hierarchical temporal memory and sparse distributed representations enhances robustness and accuracy.
• This approach advances the field of autonomous robot navigation, particularly for applications requiring persistent and reliable localization.