Cooperation of networked multi robot systems using control theory and machine learning

点击翻译按钮获取中文摘要

Autonomous vehicles interact among themselves using specialized sensing, perception, communication, and actuation suites tailored to enhance their navigation capabilities in very challenging environments. Nowadays, the uses of autonomous vehicles, robots, and agents is expanding to diverse areas such as self-driving, military missions, agriculture, and production lines, where collaborative activities and interactions with other robots in assisting humans is necessary. However, current methods used for interactions are based on either control theory concepts or statistical machine learning, making the relationship with humans complicated as humans do not perceive situations through probabilistic calculations. Even though these robots can explore unstructured environments, locate themselves, and map their surroundings, they still lack a degree of predictability derived from human understanding of causality. Unfortunately, this investigation is not very advanced in robotics, especially when considering multiple robots. Therefore, the use of robots outside of manufacturing lines and in unstructured settings like homes, schools, and hospitals is still rare, for the inability of robot to be easily understood by humans limit their usability. To perform such activities, it is not sufficient for robots to communicate with humans by using displays or voice synthesizers, but it is necessary that humans intuitively understand what robots do and why. This requires the construction of causal models that are like those used by humans. In other applications, such as the military, these characteristics are also important, for humans must be able to trust that the decisions made by autonomous vehicles and robots follow a cause-and-effect rationale. Impressive as it is, it seems to be very unlikely that this can be achieved using the advances made in statistical machine learning. This research program aims at developing novel techniques based on causal inference systems and machine learning for efficient control of networks of autonomous vehicles and robots. These techniques will take advantage of the communication network established by the vehicles but will be based on the identification of causal structures in the data and in the interactions. The research program also contemplates the use of modern machine learning architectures not only based on data, but also on models that can be understood by humans. New theoretical results will be generated with a range of applications in mind, from human-robot interactions to military applications, with close research collaboration with industry. Analyses of social interactions and dilemmas will be considered as well as classical environments including multiple vehicles pursuing others and games played by children such as capture the flag. A total of six graduate and five undergraduate students will be trained with the support of this grant over the next five years, contributing to Canada's highly qualified personnel.