Citation: Fangwen Yu. Brain Inspired SLAM System for 3D Environments [D]. Wuhan, China: China University of Geosciences, 2020: 1-218.
Abstract: With the development and widespread applications of land, sea, air and space-based robots, the autonomy, robustness and intelligence of the robot system may face new grand challenges. How to make robots truly autonomous has become a popular research topic. As one of the key technologies to realize robot autonomy, three-dimensional (3D) localization and mapping (SLAM) still faces many challenges in practical applications under unknown, unpredictable, complex, 3D environments. Furthermore, there are strictly limitations on the robot, such as the on-board sensors, computing, power consumption, storage, weight, size, etc. In addition, sensor signals are susceptible to interference and the system cannot rely on external signals such as GPS. Therefore, the conventional SLAM technology has become a bottleneck restricting mobile robot applications. How to develop a new intelligent 3D SLAM technology with characteristics of extremely low power consumption, high efficiency, high robustness and high-level intelligence has become a major problem that need to be overcome urgently. However, in the nature, humans and animals possess superlative navigation capabilities, robustly navigating over large, 3D environments, leveraging an internal neural representation of 3D space combined with external multi-sensory cues and 3D self-motion cues. For example, bats can easily navigate in complex, dynamic, 3D environments efficiently and robustly without acquiring high-precision maps only through the senses of the eyes, ears, and small brains. So how does the brain perform 3D navigation intelligently? In recent years, neuroscientists have gradually discovered internal "3D maps" and "3D compasses" in the brain, which are composed of 3D place cells, 3D head-direction cells, and 3D grid cells, etc., and gradually revealed the mysteries of 3D navigation in the brain. Therefore, can we learn from the 3D navigational neural mechanisms of the biological brain with characteristics of intelligence, efficiency, and robustness to overcome the major issues of intelligent 3D SLAM for mobile robots?
This dissertation presents a novel brain-inspired SLAM system for 3D environments named NeuroSLAM, based upon computational models of 3D place cells, 3D grid cells, 3D head direction cells, integrated with a 3D vision perception that provides external visual cues and 3D self-motion cues. The conjunctive 3D place cell and 3D head direction cell models can represent 4DoF (degrees of freedom), 5DoF or 6DoF pose according to the robot’s motion modes adaptively. The conjunctive 3D grid cell models can perform 3D path integration based on 3D visual odometry and perform 3D pose calibration based on local view cues. The 3D grid cell network dynamics drives to perform 3D path integration including 3D translational integration and 3D rotational integration in 3D manifold space. The neural network activities of local view cells, 3D place cells and 3D head direction cells drive the creation of a 3D graphical spatial experience map in a real-time, enabling relocalization and loop closure through sequences of familiar local visual cues. A 3D experience map relaxation algorithm is also used to correct cumulative errors in 3D path integration after loop closure. Using both synthetic and real-world datasets comprising complex, indoor and outdoor 3D environments, the results demonstrate NeuroSLAM consistently producing topologically correct three-dimensional maps which can efficiently support intelligent and autonomous navigation for mobile robots in 3D environments.
The main contributions of this dissertation are as follows:
Firstly, a set of computational models consisting of 3D place cells, 3D head direction cells is proposed for representing 3D pose. The computational models can represent and maintain 4DoF, 5DoF or 6DoF pose according to the motion mode of the robot adaptively.
Secondly, a set of 3D grid cell models is proposed for performing 3D path integration based on 3D visual odometry and performing 3D pose calibration based on local sensory cues. The neural network dynamics drives to perform 3D path integration including 3D translational integration and 3D rotational integration in 3D manifold space, and perform 3D pose calibration based on local sensory cues. This novel approach to 3D path integration can support 3D state estimation using a non-linear form in unknown, unstructured and unpredictable environments with non-Gaussian noise.
Thirdly, a novel encoding method of 3D graphical experience map is designed by combining local view cells, 3D place cells, 3D head direction cells and 3D visual odometry, which can not only encode a rich spatial-view place experiences, but also maintain some biological plausibility. This encoding method can also reduce false positives and repeatedly correct loop closure even when facing accumulative odometry error.
Fourthly, a light-weight 3D vision perception module is designed for estimating 3D self-motion and local sensory cues via low-cost camera sensors. Using a set of consecutive images with very low resolution, this approach can estimate 3D self-motion and sensory cues efficiently and robustly, which can also overcome the problem of environment change, such as light change, blur, etc.
Finally, an open source software system of brain-inspired 3D SLAM is developed. The system is designed based on some advanced software design principles and patterns, such as object-oriented approach, single responsibility principle, interface segregation principle, dependency inversion principle, stable dependencies principle, etc., which is capable of reliability, maintainability, scalability, extensibility, reusability.
Keywords: Brain-inspired 3D SLAM, 3D pose representation, 3D path integration, 3D experience map, 3D vision perception, 3D place cells, 3D head-direction cells, 3D grid cells, multi-dimensional continuous attractor network