Towards a microgravity sim-to-real training environment for robotic systems in low earth orbit

Human exploration towards the outer solar system necessitates higher autonomy in robotic systems for crew assistance. The development of reinforcement learning-based robotic control systems requires simulation-to-real environments reflected of those in microgravity habitats such as the International Space Station. Using NVIDIA's Isaac Lab, an open-source framework for robot training, this article outlines progress towards training space robots in microgravity environments. Isaac Lab allows for the import of custom robot models for a variety of applications and demonstrations in common workflows in robotics research. Such applications include NASA's Astrobee and Robonaut2. By configuring the environment to emulate microgravity habitats, neural network-based robotic controllers can be trained under realistic space dynamics. The framework's GPU acceleration enables efficient training of complex policies, expediting the development of autonomous robotic systems capable of performing tasks such as manipulation, locomotion, and navigation. This integration of high-fidelity simulation and advanced learning algorithms enables simulation-to-real reinforcement learning as a pivotal tool in the advancement of space robotic research.