Tactile regrasp of objects with dynamic center-of-mass

• Main Author: Than, Duc Huy
• Other Authors: Lin Zhiping
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2023
• Subjects: Engineering::Electrical and electronic engineering::Control and instrumentation::Robotics
• Online Access: https://hdl.handle.net/10356/167487
• Institution: Nanyang Technological University

Many household objects have a container-like shape, with contents that can move inside. When the contents are heavier than the container, their movement can cause the object’s center of mass (CoM) to shift. If a robot grasps the object far from the CoM, this can induce rotation and create a moving target that is difficult for a control policy to track. This project proposes a regrasp policy that utilizes the Gelsight tactile sensor to train a stability classifier and value-based DQN agent. The goal is to enable the robot to grasp objects with dynamic CoM with as few regrasps as possible. The proposed approach employs offline Reinforcement Learning (RL) to achieve sample efficiency during the data collection and training process. Various design choices and hyperparameters of the RL agents are explored and the best-performing agent exhibits high adaptability, due to its ability to adjust the step size in each regrasp attempt.