Real-Time State Estimation of a Two-Wheeled Inverted Pendulum Robot for Motion and Navigation Control

Robotic systems appear in a wide varieaty of fields including manufactury, space exploration, laboratory research and surgery. The spread of robotic technology in this wide range of applications has been motivated by their capability to perform jobs more accurate, more reliable and, last but not least, the possibility of eliminating harmful task for humans. The complex environments where robotic systems are employed require them to be able to adapt an learn. To develop different motion learning strategies and evaluate them in real-world applications, a robotic testbed has been developed capable of performing dynamic maneuvers. An autonomous sytem relies to make decisions and compute new actions. Therefore, a key aspect in making a robot really autonomous is the state estimation problem. The investigated robotic system is a Two-wheeled inverted pendulum robot (TWIPR), shown in the figure below, with instable and complex dynamics. In order to develop control strategies to stabilize the system and enable motion control of the orientation, velocity and position one needs a reliable, accurate and real-time capable estimation of the system's motion states. In order to estimate all dynamic states of the system, the robotic testbed is equipped with different sensors, such as inertial measurement units (IMUs), odometry sensors and an optical motion capture system. The estimation problem can be separated into the rotational and the translational states. Tracking of the robot's orientation (more specifically it's inclination) is a key aspect for the state-feedback controller that stabilizes the system. We therefore implement a fast, accurate and magnetometer-free estimation algorithm based on an extended Kalman filter (EKF) to track the robot's inclination to enable a high-frequency attitude controller. The translational states such as the robot's velocity and position are used for motion and navigational control in upcoming learning experiments. The multitude of different sensors providing information on the motion states and the inherent challenges of localization tasks and non-linearities in such require the use of a particle filter-based approach. This type of filter makes no assumptions on the shape of uncertainties of the different information sources and can track different hypothesis of the motion states simultaneously. This makes such a filter ideal to deal with competing information, high uncertainties, non-linearities and temporal loss of information sources. This last scenario is often encountered if the main source of localization information is prone to occlusion such at is the case with the optical motion capture system. The estimation algorithms are tested in simulation and experiments. The main objective of the simulational studies is to identify optimal filter parameters for different scenarios and to identify the method's sensitivity to different sources of error and uncertainties. For this, simulational models for the sensors and the robot are identified and implemented. The experimental evaluation is performed with a TWIPR to show that the methods work in a real-world environment and to verify the identified filter parameters. For this, several experiments are performed with different scenarios to test all methods seperately, possible sources of error and more complex scenario that simulate the robot's objectives in learning experiments.