Machine Learning and Heuristic Hybrid Control System for Complex Manipulation Automation in Space: Flexible Object Handling for Intra-Vehicular Task Assistance

The application of robotic assistance in astronaut missions is highly anticipated to enhance the efficiency and effectiveness of human spaceflight endeavors, given the limited and valuable time of astronauts. One of the major challenge in robotic crew assistance is a dexterous manipulation capability to share handling interfaces with astronauts. We focus on the flexible objects manipulation, which is heavily used in human spaceflight, yet challenging for robotic manipulation - especially in micro-gravity conditions. For this reason, we bring in deep learning approaches integrated with various robotic hardware including controllers with a strong MLOps system design in mind, and developed a system namely Predictive Learning System (PLS). We chose one of the most fundamental daily task, yet remains as a challenging task for robots on human spaceflight mission, which is to handle flexible objects. In particular, we focused on the cargo-handling tasks - opening and closing a cargo transfer bag (CTB). The results show our system is able to reach on averages 86.0% and 81.7% success rates from known & unknown target positions respectively for opening and closing the CTB. Furthermore, we have examined the application of such machine learning approach to be independent of gravitational conditions, enabling ground collected data to be applied to a different gravity conditions.