RI-IWEAR Research Seminar XI

Autonomous Robotic Capture of Non-Cooperative Space Debris

by Reinforcement Learning based Tactile Feedback Control

Since the first launch of Sputnik in 1957, the population of space debris has grown exponentially and threatened to satellite safety. To mitigate this risk, it is advised to remove five large debris pieces from space annually. Among all suggested removal methods, using robotic manipulators for capturing and removing debris is the most promising. The robotic manipulator, combined with computer vision and tactile sensors, is capable of synchronizing with the translational and angular velocities of the target, grasping and directing it towards Earth's atmosphere. In this work, we introduce a model-free machine-learning approach for a robotic system including a 6DOF manipulator and a 3-finger Robotiq gripper. Within the framework of off-policy reinforcement learning, the tasks of target pose estimation, robotic path planning, and grasping control are conducted using Deep Deterministic Policy Gradient and soft Actor-Critic algorithms. This model is exclusively trained in simulation using the Pybullet environment, eliminating the necessity for experiments or prior task knowledge. The simulation-based training allows us to delve into a broad spectrum of scenarios and conditions, which in turn enables the agent to master a diverse and reliable grasping policy. The anticipated outcome of this approach includes improved performance in real-world grasping tasks, for example, capturing free-floating objects in unstructured environments or objects with uncertain positions and orientations. Finally, we will introduce the development of a hardware-in-the-loop 3D microgravity testbed with dual robotic manipulators and active gravity compensation.