Reinforcement Learning for Robotic Rock Grasp Learning in Off-Earth Space Environments
Term
4. semester
Education
Publication year
2022
Submitted on
2022-06-01
Abstract
The time delay in communications between Earth and Mars pose a challenge for the control of rovers on Mars, for tasks such as rock grasping. This paper investigates the application of Deep Reinforcement Learning (DRL) as a method for robotic grasp learning of Martian rocks. NVIDIA Isaac Gym employed to simulate a randomly generated Mars-like environment, populated with randomly generated rocks. Proximal Policy Optimisation (PPO) is used to train a Franka Emika Panda robotic manipulator in the task of picking up the randomly generated rocks. As a Mars environment can have large changes in terrain height, joint control of the robot manipulator is implemented to enable 6-DOF grasping of rock samples. The proposed method show a 91.51% success rate, grasping randomised rocks ranging from 40-400 cm3 in size. The randomness of the environment is found to create a robust agent, while laying a solid foundation for sim-to-real in the future. This paper presents a novel approach for 6-DOF object grasping, using joint control and domain randomisation.
Documents