Enhancing Robotic Grasping of Free-Floating Targets with Soft Actor-Critic Algorithm and Tactile Sensors: a Focus on the Pre-Grasp Stage

Abstract

Capturing space debris forms the foundation for both on-orbit servicing and the removal of space debris. The challenge in this process arises from the absence of stable anchoring points and the unpredictability of motion parameters, rendering conventional spacecraft manipulator techniques unsuitable for space debris capture. In this research, we have created a method rooted in deep reinforcement learning to tackle the intricate challenge of robotic gripping, informed by tactile sensor data. Rather than manually crafting features, deep learning affords us the ability to simplify the task, fostering a learning environment for the robot to adapt grasping strategies through a process of experimentation. Our procedure implements an off-policy reinforcement learning architecture, utilizing the advanced algorithms to optimize the robotic gripper's capacity to handle objects not fixed in space, focusing on enhancing the fine grasp success rate. To efficiently master the gripping task, we have formulated a distinct reward function that delivers precise and meaningful feedback to the learning agent. Our system is trained completely within a simulated setting provided by the PyBullet environment, eliminating the need for demonstrations or pre-existing task knowledge. For this study, we examine a scenario where a Robotiq 3-Finger gripper is required to navigate toward a floating object, pursue it, and ultimately secure it. By conducting agent training within a simulated environment, we can test a variety of situations and conditions, equipping the agent with a resilient and adaptable gripping policy.