Enhancing Quadruped Robot Design With Intelligent Physics-Informed Neural Network-Assisted Dynamic State Estimation and Active Spine Integration
Loading...
Authors
Liu, Yuqing
Issue Date
2024
Type
Thesis
Language
Keywords
Alternative Title
Abstract
This dissertation represents my master's work on quadrupedal robot systems at the Autonomous System Laboratory, University of Nevada, Reno. It primarily focuses on the intelligent design of quadruped robots and unmanned vehicles. The intelligence in these designs stems from advanced state estimation techniques and the integration of an active spine, combining constraints from physical models with insights from learning-based studies. An advanced Robot State Estimation (RSE) methodology is introduced, which uses a combination of a Physics-Informed Neural Network (PINN) and an Unscented Kalman Filter (UKF) with proprioceptive sensory data to enhance state estimation accuracy. This approach effectively calibrates the Inertial Measurement Unit (IMU), mitigates IMU drift through constraints applied via Ordinary Differential Equations, and eliminates the need for external contact sensors by identifying terrain interactions, improving odometry, and operational reliability in real-world scenarios. Next, the focus is on enhancing the physical limitations of traditional robotics platforms. To achieve this, we utilize a dynamic spine to enhance flexibility and absorb impacts, which necessitates reevaluating the robot's dynamic model. To manage these complexities, the physical model estimation is enhanced, and Reservoir Computing is employed for real-time adaptive control. This significantly improves robotic mobility and stability in challenging environments.