Robotic Pushing Manipulation through Online Vision-Based Estimation of Unknown Model Parameters

Abstract

Robotic pushing is a key non-prehensile manipulation skill for handling objects that are difficult to grasp, such as thin plates. The main objective of robotic pushing is to manipulate an object along a given nominal trajectory. This requires a comprehensive understanding of the pushing dynamics model. The dynamics of pushed objects are conventionally approximated using the ellipsoidal limit surface model, which requires prior knowledge of the object’s physical properties as model parameters. However, the lack of knowledge re- garding the model parameters of these objects is a significant challenge for robotic pushing in practical applications. This issue highlights the importance of developing an online estimated pushing model for real-world applications. In this thesis, A vision-based online model estimation is proposed to push novel objects along a given nominal trajectory without prior knowledge of model parameters, such as friction coefficients and the position of the center of friction (CoF). To estimate unknown model parameters using only the vision system, a velocity-motion pushing dynamics model was derived to describe the induced motion of the object caused by the robot’s end-effector velocity. To capture the local behavior of pushing objects online, a moving-window Un- scented Kalman Filter (UKF) is used as a parameter estimator. The robotic pushing framework is constructed by combining the proposed vision-based on- line estimated model with the model predictive control-based pushing strategy. In real-robot experiments, the robotic pushing framework with the proposed vision-based online estimated model demonstrated a higher ability to follow the given nominal trajectory with greater accuracy compared to the conventional fixed model approach. Additionally, an autonomous dishware collection robot demonstration was conducted to demonstrate the applicability of the robotic pushing framework using the proposed vision-based online estimated model in various real-world applications.