Multimodal fusion for navigation of autonomous mobile robots using a deep learning method

• Main Author: Lee, Anthony Wen Hao
• Other Authors: Moon Seung Ki
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2023
• Subjects: Engineering::Computer science and engineering::Computing methodologies::Artificial intelligence; Engineering::Industrial engineering::Automation; Engineering::Electrical and electronic engineering::Control and instrumentation::Robotics; Engineering::Mechanical engineering
• Online Access: https://hdl.handle.net/10356/167600
• Institution: Nanyang Technological University

Differential-drive robots have become increasingly prevalent within human society and have become an essential tool in various industries and fields, such as manufacturing, healthcare, and entertainment. This is because these robots are usually cheaper than more advanced robots such as, omnidirectional-drive robots and are easier to control. As differential-drive robots become more ubiquitous, there is a growing need for better collision-avoidance in navigating robots. Although differential-drive robots possess collision avoidance capabilities, they usually face challenges in navigating human environments that are more crowded, thus there is a demand for better capabilities in differential-drive robots. Many studies have been done to explore new approaches such as using advanced sensor technologies and machine learning algorithms, which aims to improve the accuracy and efficiency of these collision-avoidance methods for robots. However, most existing, or newly collision-avoidance methods developed for holonomic robots are not always effective in ensuring the smooth operations of differential-drive robots. Even if it is operational, it does not guarantee that the navigation will be efficient and safe, which is crucial in human environments. Thus, more research is needed to develop better navigating solutions in differential-drive robots, since it is widely used in human societies. In this research, there is exploration on various type of navigation methods that aims to address social interactions between humans and robots. The proposed research is conducted to improve on one of the selected decentralized collision-avoidance method that is for holonomic robots, so that it can be applicable for differential-drive robots. The improvement is through the addition of mathematical equations to the chosen collision-avoidance method’s policy which allows the differential-drive robot to ii complete navigation tasks while moving smoothly. The mathematical equations are adopted from another research that allows differential-drive robots to move like a holonomic one. A virtual differential-drive robot’s states are inputted into the policy for holonomic robots. Next the policy will undergo simulations with its own simulated omnidirectional-drive robot. Subsequently the policy will return outputs on the future states of the differential-drive robots for its next time steps. Experiments is then executed, and the results of the proposed method will be compared to the original method for holonomic robots. It is projected that the differential drive mobile robot would be performing similarly to the holonomic robots, when trained under the same policy. Finally, visualization is used to validate the movement of the robot when its moving, with the new robot navigation method developed in this report.