Grant: Discovery Grants Program - Individual

Quadruped robots are increasingly used for a wide range of applications. They have remarkable agile capabilities; however, most are still remotely piloted because they lack a high level of self-awareness. Reliable state estimation (SE) is a significant aspect for the deployment of quadrupeds because most locomotion planners and control strategies rely on an accurate pose and velocity estimate of robots. The long-term vision of this research program is to improve the autonomy of quadruped robots by developing novel perception algorithms. More specifically, we will enhance the robot's position and velocity estimates in dynamic environments. We will do this with two main approaches. The first approach is to improve mapping and visual-inertial sensor fusion by including both robot and environmental dynamics in SLAM. Current literature is focused on adding exteroceptive sensors to provide a non-drifting pose for navigation. Furthermore, it often focuses on wheeled robots in well-structured environments. Some drawbacks of these approaches are the frequencies and delays in the sensors and the assumption of static environments. These works can be improved by including robot and environmental dynamics into the SLAM formulation. They are often not included due to the theoretical challenge of incorporating the nonlinear dynamics into the problem. Additionally, the dynamics of the environment are entirely unknown. Other outstanding challenges include overly optimistic filters caused by inconsistencies in the covariance estimates and SLAM algorithms can be too slow for accurate foot path planning. The second approach uses a ground-based robot and an aerial robot to perform coordinated SE. One disadvantage of a camera on a quadruped is the limited field of view. A flying drone can easily overcome this limitation by flying above the ground-based robot. An outstanding challenging problem for collaborative visual SE is that most existing approaches to formation control assume that every robot has a global navigation system and all following robots have the reference trajectory of the leader. However, this is not practical. Using computer vision (CV) is critical to improving autonomy. Other issues, such as robustness to disturbances, time delays, and sensor noise, must also be considered and studied in a coordinated control problem context. This nonlinear control problem is interesting, critical, and relatively unexplored. This research will contribute to new theories and bridge the gap between theory and actual implementation. The complexity of visual SE and coordinated control requires a deep analytic approach for algorithm performance, while the practical implementation will demonstrate the research's use to society. This research proposes to advance the state-of-the-art by combining CV, the robot's dynamics, and nonlinear control to make robots more autonomous. Additionally, this research will contribute to the training of HQPs for future robotic research.