Robotic manipulation of objects through physical simulation

Abstract

In recent years, robotics has profoundly impacted automated industries, with robotic manipulators revolutionizing efficiency and allowing humans to dedicate time to more complex tasks. However, traditional robotic systems are programmed for predictable environments, creating a growing interest in replicating human adaptability in these systems — especially in scenarios where object manipulation is necessary. Humans form beliefs from neural processing of sensory information, which allow us to predict and anticipate the outcome of a situation by assigning higher probabilities to the case that corresponds the most with our past experience and sensations. This aligns exactly with the operational principles of the Bayes Filter, making it a promising method to implement in robotic manipulation to improve the reliability of robots in performing tasks within uncontrolled environments. Inspired by the Bayesian Brain theory, this dissertation aims to address this challenge by developing a system that utilizes Bayesian particle filtering to estimate physical properties of an unknown object, such as the mass and the center of mass, through manipulation. In this approach, we generate a range of potential simulation scenarios, where each instance corresponds to a particle representing a possible state of the object within a robot-object interaction. As the robot interacts with the object, it gathers sensory measurements, which are used to iteratively update the particles through the filtering process, until the particles converge toward a single, accurate estimation of the object’s properties. The evaluation performed for the solution demonstrates the proposed method’s effectiveness in enhancing a manipulator’s adaptability to an unpredictable environment. By integrating the principles of probabilistic robotics into manipulation, this research contributes to the development of more capable robotic manipulation strategies in real-world applications, and confirms the viability of Bayesian filtering as a technique for robotic adaptability.